Formulation Optimization AI: Data Design, PoC and RFP for Materials Informatics in Thai Plants

Interest in “formulation optimization AI”, which searches for formulations close to target properties with fewer trial batches, is growing in the development departments of materials and chemical manufacturers. Paints, inks, adhesives, sealants, rubber compounds, resin compounds, lubricants, cleaning agents. Technology development departments in Thailand receive a steady stream of property requirements from customers, cost reduction requests, switches to alternative raw materials and replacements of regulated substances. Meanwhile, formulation trials are mostly run by changing one condition at a time based on the experience of veteran engineers, and the number of trials and the time available cannot keep up. This is where formulation optimization AI comes up for consideration: the introduction of so-called materials informatics (MI), which uses machine learning and experimental design methods such as Bayesian optimization.

This article makes one argument. Whether formulation optimization AI succeeds depends not on the choice of algorithm, but on whether you decided the following three things first. (1) Defining goals and constraints: which properties, in what ranges, in what order of priority, and how to bound raw material upper and lower limits, the 100% total, regulated substances and the cost ceiling. (2) Structuring experimental data: recording formulation ratios, raw material lots, process conditions, evaluation conditions and even failed experiments in the same format, and measuring the “usable share” of past data first. (3) Running the experiment loop: who approves the next trial batch proposed by AI, after how many rounds to stop, and when results are fed back. AI proposes “the formulation to try next.” But it does not decide “whether that formulation may be made (safety, regulations, equipment)” or “where to stop development.”

Below, so that heads of technology development and R&D, quality assurance staff and plant managers at Japanese-affiliated and local materials and chemical manufacturers in Thailand (and ASEAN), as well as research laboratories and technology planning departments at head offices in Japan, can assemble the material they need for internal decisions, we go through trends, the differences between methods, why trial counts balloon, how to set goals and constraints, data inventory, how to run the experiment loop, choosing between products and in-house builds, integration with existing systems, cautions when proceeding in Thailand, a model calculation of costs and benefits, RFP items and PoC, in that order. The model plant’s project counts, trial counts, unit costs, amounts and payback periods in this article are all placeholder figures set for this article. They are not industry averages, survey values, or figures from vendors or papers.

Why AI for formulation optimization now

What materials informatics refers to

The term materials informatics covers different scopes depending on who uses it. Uncountable, which provides an MI platform, defines MI as “the practice of centralizing and structuring materials data, then applying data analytics and machine learning to accelerate development,” and lists as its foundation the centralization of data, structuring that makes it searchable by content, and connecting R&D with quality control and product lifecycle management (a vendor definition). Hitachi High-Tech describes MI as something that “uses AI technology on experimental data accumulated in the past to predict physical properties and derive optimal conditions for manufacturing conditions and material formulation ratios” (company release).

Both descriptions have in common that they place “getting the data in order” before the algorithm. This article also treats formulation optimization AI as “a mechanism that structures formulation ratio and result data, predicts properties with machine learning, and proposes the formulation to try next with Bayesian optimization or similar methods.”

Public projects have shaped the field

In Japan, the Materials Research by Information Integration Initiative (MI2I) of the National Institute for Materials Science (NIMS) started in July 2015 as part of the Japan Science and Technology Agency (JST) “Innovation Hub Construction Support Program” and ended at the end of March 2020. It worked on data-driven materials science research and is said to have formed a community for the new field of materials informatics, and activities have continued at NIMS and elsewhere using that network after it ended. The NIMS Data Platform provides “DICE”, a cloud platform that accumulates materials data from diverse sources such as experiments, measurements, calculations and literature in a form that is easy to reuse (the page does not state whether, or under what conditions, companies can use it).

In the formulation domain, NEDO (New Energy and Industrial Technology Development Organization), ADMAT, AIST (National Institute of Advanced Industrial Science and Technology) and Showa Denko (now Resonac) announced on April 13, 2020 that, using AI in a national project on foundational technology for ultra-high-speed development of advanced materials, they had “reduced the number of experiments to 1/25 or less of the conventional approach” in developing a flexible transparent film. The conditions were that 27 types of film were initially made and the AI was trained on the measured data, and the AI predicted three properties: converted transmittance, breaking stress and elongation. The property values of the films with AI-predicted formulations reportedly exceeded those of the initial 27 types and were better than the 25 types made by skilled researchers in a comparison experiment. However, this is the result for a specific material and a specific theme, and the details of how “1/25” was calculated cannot be read from the announcement page. It does not mean the same ratio will be achieved in your own formulation development.

Company announcements are increasing (but the conditions vary)

Company announcements have followed. All of them are announcements by the companies themselves or by the tool providers, and none have undergone third-party verification. They must always be read together with the comparison baseline and the scope.

• Mitsui Chemicals and Hitachi, Ltd. (June 28, 2021): Announced the start of a verification trial applying Hitachi’s AI-based MI technology to new materials development at Mitsui Chemicals. In a verification using past development data on organic materials provided by Mitsui Chemicals, the number of experimental trials was reportedly “reduced to about 1/4 compared with conventional MI.” The comparison baseline is “conventional MI,” not human trial and error, and it is a verification using past data.
• Hitachi High-Tech and NOF Metal Coatings (August 19, 2025): Announced that at NOF Metal Coatings, which works on high-performance thin-film coatings such as anti-corrosion materials, “in research and development work on a specific theme, a reduction of 50% or more in the number of experiments and time was achieved compared with before MI was used.” It states that this was carried out with hands-on support from Hitachi High-Tech data scientists and others, as the company had no MI or data analysis specialists in-house.
• Sika (November 27, 2025): Sika, a construction chemicals manufacturer, states in its investor and media materials, for the product example “Nuage”, “75% fewer experiments with machine learning” and “time to market reduced by more than 50%.” These are values for a specific product example of the company, published in Sika’s own materials.
• Advanced Polymer Coatings and Albert Invent (June 23, 2026): A US manufacturer of anti-corrosion polymer coatings announced that it will introduce Albert Invent’s “AI-native OS.” The release states that AI-driven experimental design and formulation optimization have been shown to cut the company’s development time by 50% (a release by the tool provider).
• SCG Chemicals: A case in which SCG Chemicals (SCGC), one of the largest integrated petrochemical companies in ASEAN, introduced Uncountable’s R&D platform and replaced Excel- and Python-based tools has been published as an Uncountable case study (details below).

When numbers are lined up like this, it is tempting to read them as “AI halves trial batches,” but the comparison baseline (humans or conventional MI), the verification method (actual trials or reproduction with past data) and the scope (a specific theme or company-wide) all differ. The only way to know your own reduction rate is through retrospective verification on your own past projects and a PoC.

Why “trials cannot keep up”

The background to formulation trials falling behind is similar in every plant. Property requirements from customers (automotive, electrical and electronics, packaging and so on) become more detailed, while cost reductions are demanded at the same time. When localizing raw material procurement or switching to alternative raw materials, the whole formulation has to be readjusted each time one raw material changes. Replacing regulated substances comes with fixed deadlines. A Citrine white paper claims that, for replacing problematic ingredients such as PFAS, a model trained on a base formulation can be applied across a product family to propose alternatives (vendor claim). Regardless of how far this applies, the structure in which “change one thing at a time” trials cannot keep up as replacement projects increase is a concern shared by many development departments.

What formulation optimization AI (materials informatics) is: DOE, machine learning, Bayesian optimization and generative models

What is called formulation optimization AI mixes methods of different character. When discussing internally, agreeing on which method you are talking about also makes vendor explanations easier to read.

This table summarizes the general character of each method. For generative models, we have not confirmed concrete results in formulation optimization within the scope of this article’s research, so we limit ourselves to introducing where the method sits.

The typical flow of Bayesian optimization, as illustrated on the official site of the open-source BoTorch, is an iteration of fitting a Gaussian process model, constructing an acquisition function (such as Expected Improvement), and optimizing that acquisition function to decide the next candidate.

Formulations have a distinctive constraint: “each component’s ratio is 0 or more, and the total is 1 (100%).” Research on handling this condition mathematically is also progressing. An arXiv preprint published in March 2026 (not peer reviewed) proposes “α-GaBO”, a Bayesian optimization algorithm for spaces with a sum-to-1 constraint (the probability simplex), and claims in its abstract that it showed higher performance than methods on constrained Euclidean space on benchmark functions and real problems including mixtures of components. The abstract does not state specific material systems or numbers of experiments.

There is also research on actual formulations. A paper published in Cell Reports Physical Science combined a machine learning classification algorithm with the multi-objective optimization algorithm TSEMO to design liquid formulation products for which no physical model exists, and ran semi-automated experiments with a robot. Its abstract states that it found 9 recipes meeting criteria set by the customer within 15 working days and “outperformed human intuition” on target performance. This is a result in a robotic experimental environment, and within this article’s research we could not check the full text for comparison conditions such as the number of experiments on the human side.

As a practical vendor explanation, a Citrine white paper for paints and coatings shows how to proceed from a state with no data as follows (vendor claim). First, “prime” the AI model with initial experiments that resemble a DOE matrix but are narrowed down to cover the search space with the minimum number of experiments. After that, as “sequential learning”, repeat a cycle in which the AI proposes about 5 experiments at a time, they are carried out, the results are entered, and the model is retrained to propose the next ones. The company claims this methodology requires fewer experiments than trial and error or DOE.

In other words, DOE and Bayesian optimization are not opposed. Designing the initial experiments in a DOE-like way and proceeding sequentially with Bayesian optimization from there is also the approach shown in the Citrine explanation above.

Five reasons trial batches pile up in formulation development

Before introducing AI, separating out why trial counts balloon makes it clear which parts AI can reduce and which parts it cannot.

The rightmost column of this table is what should be decided before introduction. In particular, for (3) to (5), no matter how good the AI algorithm is, it will not work if the data has not been kept. Efforts to keep the basis of veteran engineers’ judgments as data also overlap with the ways of capturing tacit knowledge covered in our article on skill transfer AI.

Note that themes optimizing process conditions (injection molding temperature, pressure, holding time and so on) are easier to organize when treated as a problem separate from optimizing formulation ratios. For process condition optimization, see our article on injection molding parameter optimization AI, and for yield improvement in mass production, see our article on yield improvement AI. This article focuses on deciding “material formulation ratios (recipes)” at the development stage.

“Goals and constraints” to decide before introduction: target variables, acceptable ranges, upper and lower limits, 100% total, regulated substances, cost

What deserves the most time in preparing for formulation optimization AI is defining goals and constraints. If this remains vague, the AI will keep proposing formulations that look plausible but cannot be used.

The figure above views formulation optimization AI as four layers. At the top are “goals and constraints”, below them “data”, then “experiment loop”, and finally “measurement and acceptance”. If an upper layer is not decided, the layers below cannot be designed. The Citrine white paper also lists three places where knowledge is stored: (1) data uploaded to the system, (2) AI models that represent the relationship between inputs and outputs, and (3) the “search space”, that is, a description of formulation constraints such as usable raw materials and mixing parameters (vendor claim). (3) corresponds to the goals and constraints layer described here.

(1) Narrow down the target theme with four conditions

The choice of the first theme largely determines the results of formulation optimization AI. It is safest to start with a theme that meets the following four conditions.

1. Property targets can be measured numerically (viscosity, hardness, tensile strength, adhesion grade and so on. Themes evaluated only by sensory evaluation come later)
2. Each trial is costly in money and time (expensive raw materials, evaluation tests that take days, external testing required and so on)
3. Variables fit within around 10 (counting raw material types and process conditions together, there are not too many variables to move)
4. The evaluation method is stable (the variability when the same sample is re-measured is sufficiently smaller than the improvement you aim for)

(2) Target variables and acceptable ranges

Write target variables not only as “what to maximize” but also as “what range is acceptable.” For example, decide target values, acceptable ranges and priorities, such as “viscosity in the range of X to Y, hardness of Z or more, raw material cost of W or less.” When multiple properties conflict (making it harder reduces elongation, for example), decide in advance which takes priority, or whether candidate trade-offs between the two will be listed for people to choose from.

(3) Upper and lower limits per raw material and the 100% total

Decide upper and lower limits for each raw material from both technical reasons (it does not disperse, it does not cure) and business reasons (procurement volume, cost). The constraint that formulation ratios total 100% must always be enforced on the AI side. The α-GaBO research mentioned above also targets this “sum to 1” constraint.

(4) Regulated and prohibited substances

Raw materials that must not be used, components with content limits, and lists of prohibited substances specified by customers should be removed from the search space in advance. Relying only on an operation in which people notice and stop AI proposals after they are made leads to oversights. If checking against raw material SDS and regulatory information is linked to the chemical master described in our article on SDS management AI, regulatory updates become easier to reflect in the search space.

(5) What production equipment can make

Even if development finds an optimal formulation, it is meaningless if its viscosity cannot be mixed on mass production equipment or it uses raw material quantities that cannot be charged. Build in mass production constraints from the development stage, such as the capacity of kneaders and dispersers, charging order and temperature limits.

Do not add too many constraints

On the other hand, packing in too much knowledge can backfire. An arXiv preprint (not peer reviewed, November 20, 2025) by researchers at Fraunhofer IAIS and others reports a case of using Bayesian optimization to develop a plastic compound containing recycled material (four components: virgin PP, recycled plastic, filler and impact modifier). The targets were MFR of about 10 g/10min, Young’s modulus of 1500 MPa or more and impact strength of 8 kJ/m² or more, and the formulation upper limits were 30% filler and 20% modifier. When features were built from technical data sheets using expert knowledge, it became “an 11-dimensional problem,” and partly because none of the 430 past data points satisfied both constraints at the same time, the optimization got worse. Only after simplifying to just the four formulation parameters did satisfactory results appear. The authors conclude that “additional knowledge and data are beneficial only if they do not complicate the optimization objective.”

This case has an important premise. The 25 experiments (three batches of 10, 7 and 8) were an experimental campaign carried out by process engineers, and the engineers’ best MFR was 6.65 g/10min. The simplified Bayesian optimization was evaluated in a simulation that used the predicted mean of a trained model as an “oracle” (the role that returns answers), not through physical trials. The result was “10 runs satisfying the constraints, best MFR 6.13 g/10min (a value close to the skilled engineers’ result).” Bayesian optimization did not outperform the engineers, nor did it achieve the target in actual trials. What can be learned is the importance of narrowing down variables and the difficulty of starting from a state in which the data contains no example that meets the target.

Data inventory: what share of past trial data is usable

After goals and constraints comes data. In many development departments, past trial data is scattered across lab notebooks, Excel and personal PCs. The SCG Chemicals case study also lists as pre-introduction challenges that data was scattered across personal PCs, servers and various systems, that developing ML/AI models required Python skills that researchers did not have, and that polymer formulation development involved a large number of experiments (vendor claim).

Items to align

For each theme, check whether past trials can be aligned on the following items.

Measure the “usable share” first

In the inventory, sort past data into “usable as is”, “usable if corrected” and “unusable”, and calculate the shares by count. For example, count records that have formulation ratios and results but unknown evaluation conditions, records with unknown lots, and records where failures were not recorded. This share affects both the model’s initial accuracy and the cost of structuring the data.

The Citrine white paper gives an example of data silos, such as commercial raw material information in ERP, rheology measurements in LIMS and intellectual property in handwritten notebooks, and recommends first running a short-term AI project to determine which data is useful for AI models and worth digitizing (vendor claim). The order is not to digitize everything before starting, but to prioritize structuring the items needed for the target theme.

The same document says that because making and testing a single formulation sample costs “hundreds to thousands of dollars”, datasets in chemistry are small, and “small data AI” distinct from “big data AI” like ChatGPT is needed. It then lists, as measures for small data, “knowledge integration” that narrows the model with expert knowledge, “chemical featurization” that automatically generates features such as molecular weight and number of hydrogen bonds from molecular structures, and “uncertainty quantification” that calculates the probability of achieving a target (vendor claim). The Fraunhofer case above is also an example where adding knowledge actually made the optimization worse, so it is prudent to decide whether to add knowledge while verifying.

Have people review outliers and feature importance

Once the data is in, plot it and check for outliers. Outliers may be input errors. The Citrine white paper also recommends that the team review the features the AI model considers important (feature importance) and check whether the model is saying something that makes chemical sense, which helps both to confirm the model’s validity and to gain new insights (vendor claim).

For how to think about join keys between mass production quality data and development data, our article on quality data management systems is also a useful reference.

How to run the experiment loop: AI proposal, engineer approval, trial, evaluation, data feedback

Once goals, constraints and data are in place, design the experiment loop. The most important point in operating formulation optimization AI is to always put an engineer’s approval between the AI proposal and the trial.

The loop flow

1. Initial experimental plan: For themes with little past data, build initial experiments that broadly cover the search range in a DOE-like way
2. Model training: Build a model from the initial experiments and the usable past data
3. Proposal of next candidates: Propose several formulations to try next using Bayesian optimization or similar methods. Show predicted values together with their uncertainty
4. Engineer approval: Check proposed formulations from the viewpoints of safety (heat generation, hazardous gases, handling), regulations (prohibited substances, content limits), equipment (whether it can be mixed, whether it fits the trial equipment), cost and chemical validity, and approve, modify or reject them. Record the reasons for rejection as well
5. Trial and evaluation: Make the approved formulations and measure them with the agreed evaluation method
6. Data feedback: Register the results, including failures, in the same format and update the model

Days per loop and proposals per round

How many days one loop takes is largely determined by the number of days the evaluation test takes. If a test requires a week of curing, the loop takes a little over a week at minimum. How many proposals to request per round is decided by how many samples the trial equipment can make at the same time and by evaluation capacity. Citrine’s explanation gives about 5 at a time as an example (vendor claim), but this is just one example, and you should match it to your own equipment and evaluation capacity.

Decide stopping criteria first

An AI loop will continue indefinitely unless stopped. Before starting, decide stopping criteria such as the following.

• Stop once X formulations with all target properties within the acceptable ranges have been obtained
• If prediction improvement stalls after X loops, go back and review the goals, constraints and variables
• When the upper limit on trial count or duration is reached, make a decision with the best formulation obtained so far
• If engineers keep rejecting proposals (if the proposal adoption rate is low), review how the constraints are written

Where to stop development is decided by the development lead, not by AI. AI outputs “formulations that seem worth trying next,” but it does not judge “whether this is good enough.”

Keep records of approvals

The record of which proposals engineers approved, which they rejected and why is valuable data in itself. If the reason for rejection is “this combination separates,” that is knowledge that should be added to the search space as a constraint from the next round. In this way, the judgments of veteran engineers are progressively kept as data.

Build in-house or use a product: commercial MI platforms, open source, joint research

There are three broad ways to implement formulation optimization AI.

Examples of commercial MI platforms (all vendor claims)

The following introduces products for which official information could be confirmed in this article’s research, and is not a recommendation. Note that all are the vendors’ own claims, and many do not state comparison conditions.

• Citrine Informatics: Consists of Citrine DataManager (accumulating internal knowledge), Citrine VirtualLab (virtual experiments and identifying promising candidates), Citrine Catalyst (digital assistant and citation search) and professional services. Its white paper explains that it has an API but prioritizes a no-code interface, and that formulation engineers can add knowledge such as relational equations to the model themselves. The official site advertises “5x faster product development,” but no comparison conditions are stated. The white paper includes customer cases such as reaching target properties in 2 weeks instead of the usual 10 weeks for an automotive multilayer polymer coating, and reformulating a launched product to remove PFAS while maintaining performance in less than 3 months (customer names and detailed conditions are often not disclosed).
• Uncountable: On its MI explanation page, it lists results at customer companies, such as Repsol consolidating more than 10,000 data points from 20 years of formulation work into a single searchable platform and reducing experimental workload per project by 30 to 40%, and AGC Chemicals and Covestro shortening development time by up to 6 months.
• Polymerize: Consists of Polymerize Connect for data management, Polymerize Labs for prediction and optimization with more than 35 domain-specific AI models, and Polymerize One, a full service including a trial production network. Its site claims “50% fewer failed experiments within 3 months” and “3.1x ROI within 6 months.”
• Albert Invent: In the Advanced Polymer Coatings announcement mentioned above, it highlights AI-driven experimental design and formulation optimization.
• Hitachi High-Tech “Materials Development Solution”: States that it has been provided since 2021 to domestic and overseas customers, mainly materials and chemical manufacturers, and in the NOF Metal Coatings case it emphasizes hands-on support by data scientists (customer success).

Pricing: None of the official pages and releases of Citrine, Uncountable, Albert Invent, Polymerize and Hitachi High-Tech checked in this article’s research showed specific prices. Costs can only be confirmed through individual quotations.

Building with open source

• BoTorch: A “Bayesian optimization in PyTorch” library published by Meta (MIT license). It is characterized by a modular structure into which new models and acquisition functions can be plugged. However, the README recommends using BoTorch as a low-level API for implementing new algorithms for Ax (positioning Ax as the easy-to-use platform for end users), so it is positioned for researchers and advanced users.
• Ax: An adaptive experimentation platform developed by Meta. It illustrates parameter optimization that learns from trial results to improve proposals, and advertises experiment management including MySQL storage.
• Optuna: A hyperparameter optimization framework developed by Preferred Networks (MIT license). Its official collection of examples covers multi-objective optimization and constrained optimization. It is originally intended for hyperparameter search in machine learning, and applying it to formulation optimization depends on the user’s design.
• Libraries no longer maintained: The GPyOpt repository was archived on February 23, 2023, with the end of maintenance declared. scikit-optimize was also archived on February 28, 2024. Because they are sometimes introduced in older tutorials and sample code, check maintenance status when adopting a library for new work.

Even when building with open source, you need to build data management, approval screens, logs and integration with existing systems yourselves. The optimization library is only one part of the overall system. Contracting and acceptance when outsourcing development are covered in our article on outsourcing AI development.

How to choose

If a department with few projects and researchers who do not program suddenly introduces a licensed product, it may go unused. The Citrine white paper also says that “AI left out like a fruit basket will not be used,” and lists building a network of promoters (champions) and agreement from senior management as key points of change management (vendor claim). Conversely, if a department without data scientists builds in-house with open source, the system may stop as soon as the person who built it is transferred. The emphasis in Hitachi High-Tech’s NOF Metal Coatings case that the work proceeded with external hands-on support despite having no MI specialists in-house is one answer to this point.

Integration with existing systems: boundaries with ELN, LIMS, formulation masters and quality data

Run on its own, formulation optimization AI becomes “a data island for the development department only.” Decide which systems it connects to and where the boundaries are drawn.

• Electronic lab notebook (ELN): Where to keep the master record of trials. Decide whether the ELN is the master and links to the MI platform, or data is entered directly into the MI platform. Double entry does not last.
• LIMS (test and analysis management): Transcribing evaluation results by hand introduces transcription errors. Check whether measurement results can be imported automatically and whether evaluation conditions (standard, instrument, operator) can be passed along with them.
• Formulation master in the production management system: When moving a formulation decided in development to mass production, register it in the formulation master of the production management system. At this point, the raw material codes used in development and the item codes in mass production must match, or the two will not connect. Formulation master and lot management requirements for chemical plants are explained in our article on production management systems for chemical plants, and how to clean up item codes in our article on manufacturing master data management.
• Quality data: If quality problems that occurred in mass production can be fed back into development data, the same problems can be avoided in the next development. For monitoring process variation in mass production, see our article on SPC anomaly detection AI.
• Purchasing and ERP: Raw material prices and availability are in ERP. If the cost ceiling is included as a constraint, decide which point in time’s raw material unit prices to use.

The principle when setting boundaries is that “development and mass production are connected by four keys: raw material code, formulation number, trial number and lot number.”

Cautions when proceeding at plants in Thailand and ASEAN

Cases and industry-academia collaboration are emerging in ASEAN too

In Thailand, a case in which SCG Chemicals introduced Uncountable’s R&D platform has been published. The results stated in Uncountable’s case study are “45% reduction in DOE (design of experiments)”, “10 to 20% faster R&D”, “up to 3x ROI” and “50% reduction in data mapping time from multiple sources,” and the leader of the company’s Polymer Simulation team comments that “before Uncountable, we could not develop appropriate models for polymer formulations.” The introduction date is not stated, and another Uncountable page uses different wording: “20% increase in R&D output, 45% reduction in DOE workload.” All are announcements by the tool provider.

In industry-academia collaboration, Hitachi announced in an English release on June 30, 2025 that Hitachi High-Tech (Thailand) had signed a collaboration agreement with the Petroleum and Petrochemical College (PPC) of Chulalongkorn University. It aims to use informatics expertise centered on MI to advance education and research, with goals such as data science education, social implementation of research results and strengthening industry-academia collaboration (this is an announcement of future initiatives, with no duration or amount stated). Hitachi High-Tech also states that it will support MI use at NOF Metal Coatings’ overseas sites in the future (planned).

BOI R&D incentives (as described in the 2022 guide)

In the Thailand Board of Investment (BOI) “A Guide to The Board of Investment 2022”, activity category “7.11 Research and development” is classified as A1 (8-year corporate income tax exemption with no cap), and the definition of applied research explicitly includes “formula development”, product design and production process design. Conditions listed include R&D personnel costs of at least 1,500,000 baht per year (new hires) or capital investment of at least 1,000,000 baht (excluding land and working capital), and the guide states that applying for merit-based incentives can extend the corporate income tax exemption to up to 13 years, and that locating in a BOI-promoted science and technology park provides a 50% corporate income tax reduction for 5 years after the exemption period ends. In the same guide, “8.1.3 Advanced Material Technology Development” falls under Section 8 (10-year corporate income tax exemption with no cap), with technology transfer with BOI-approved educational or research institutions listed as a condition.

On June 30, 2021, the BOI also approved enhanced incentives for R&D and human resource development, stating that projects investing or spending at least 200 million baht on R&D, or 1% of total sales in the first 3 years, can receive longer tax incentives (up to 13 years).

These are descriptions from the 2022 guide and the 2021 announcement. They may have been revised since, so if you are considering applying, be sure to confirm the current conditions with the latest BOI notifications and the BOI office. This does not mean that introducing formulation optimization AI itself qualifies for incentives; it is information that an R&D activity category covering formulation development exists. The overall approach to AI development in Thailand is also covered in our article on how to start AI development in Thailand.

Data leaving the company and confidentiality

Formulation data is one of the most important confidential assets for a materials manufacturer. When using a cloud-based MI platform, check in which country’s servers the data is stored, whether the vendor uses your data to improve its models, and in what format data and models can be retrieved when the contract ends. When sharing data with the research laboratory at the head office in Japan, decide in advance with the head office information security department which formulation data is kept on the Thai side and which on the Japanese side, and who holds access rights.

Division of roles with the head office research laboratory in Japan

When pursuing formulation optimization AI at a Thai site, deciding the division of roles with the head office research laboratory in Japan reduces confusion. For example, new base formulations and method development could sit with the head office, while improved formulations for customers and localization of raw materials sit with the local site. Because data on local raw material brands and lot differences exists only locally, it is important to set up a structure in which the local development department can handle data registration and approval.

Model calculation of costs and benefits (placeholder figures)

From here on is a model calculation to show the “structure” of the estimate. All figures below are placeholder figures set for this article. They are not industry averages, research firm figures, vendor prices or results from papers. Replace them with your own figures before use.

Model plant assumptions

Assume the technology development department of one site of a Japanese-affiliated paint and adhesive manufacturer in Thailand. The currency is Thai baht (THB).

Main case (reduction rate 30%)

As a general formula, this becomes:

Annual net benefit = 2,880,000 × r − 660,000

2,880,000 is the conventional annual trial cost (15 × 24 × 8,000), and 660,000 is the sum of the additional work and the annual operating cost (60,000 + 600,000).

How the result moves when the reduction rate changes

Let us change only the reduction rate r.

The reduction rate at which the annual net benefit becomes 0 is r ≈ 22.9%, from 2,880,000 × r = 660,000. Under these placeholder assumptions, if the trial reduction falls below about 23%, trial cost reductions alone do not even cover the annual costs. Far from recovering the initial cost, the annual operating cost and additional work exceed the trial cost reductions every year. On the other hand, at a 50% reduction rate, payback takes less than 1 year. The result depends strongly on the reduction rate, and it is tougher for departments with fewer projects (because the annual operating cost is assumed to be fixed regardless of project count).

Look at cash savings and labor savings separately

Of the 8,000 THB per trial, the 6,000 THB for raw materials and evaluation tests is cash spending, but the 2,000 THB of engineer time does not become a cash saving unless the freed-up time is redirected to other development. Of the 864,000 THB trial cost reduction in the main case, cash savings are 648,000 THB (108 × 6,000) and labor savings are 216,000 THB (108 × 2,000). Looking at cash savings alone, they only slightly exceed the annual operating cost of 600,000 THB. Whether the freed-up engineer time can be redirected to the next development projects or customer support greatly changes the meaning of the investment.

What this estimate does not include

The following effects are not monetized because they are difficult to estimate.

• Earlier orders or product launches from shorter development periods
• Lower mass production costs if a formulation with a cheaper raw material mix is found
• The effect of completing regulated substance replacements within deadlines
• The effect of keeping veteran engineers’ knowledge as data

Evaluate these separately from trial cost reductions. None of them can be promised in monetary terms in advance.

What the estimate tells us

The reduction rate r must not be filled in directly with vendor claims or figures from papers. As noted above, published reduction rates are values for specific material systems, specific themes and specific comparison baselines, and there is no guarantee they apply to your company. The sound approach is to make the investment decision after measuring the reduction rate through retrospective verification on your own past projects and a PoC.

For departments with few projects, one option is to start not with a licensed product but with structuring experimental data and design of experiments (DOE), and to consider a licensed product once the number of projects grows. Structuring data is not wasted whichever approach you choose.

RFP items for formulation optimization AI

When writing an RFP (request for proposal) for product selection or outsourced development, include the following items.

What to check in the PoC: five cases to test deliberately

In the PoC, check not only whether good predictions come out with clean data, but also whether, when given adversarial conditions, the AI avoids unreasonable proposals, honestly shows its uncertainty and asks for human judgment.

Cases to test deliberately

1. Retrospective reproduction of past projects: Give the AI data from an already completed development project only up to a midpoint in its timeline, and reproduce whether the AI could have reached a formulation of the same level with fewer trials than actually used. The reduction rate measured here is the first material to replace r in the model calculation.
2. Whether constraint-violating proposals appear: After setting upper and lower limits, the 100% total and prohibited substance constraints, check that not a single proposal violates the constraints. Deliberately including a raw material containing a prohibited substance in the candidate list makes this easier to check.
3. Display of uncertainty in extrapolation regions: When targeting conditions far outside the range of past data, check whether the AI shows not only predicted values but also large uncertainty, or issues a warning.
4. Switching raw material brands: When a raw material is switched to an equivalent from another manufacturer, check how the AI handles the new brand with little data and how far existing data can be reused.
5. Evaluation variability: Enter repeat measurement values for the same formulation and check that measurement variability is not mistaken for a formulation effect.

KPIs to measure

• Trial count and duration until the target is reached (compared with the conventional approach)
• Prediction error (difference between predicted and measured values)
• Proposal adoption rate (share of proposals approved by engineers and breakdown of rejection reasons)
• Time spent on data registration (actual additional work)

Even if the PoC shows that the reduction rate did not reach the break-even level, the data structuring and constraint organization remain as a foundation for the next initiative. To be able to evaluate results honestly, decide the success criteria in writing before the PoC.

Frequently asked questions

What is materials informatics?

It is an approach that centralizes and structures materials data and uses data analytics and machine learning to accelerate materials development (based on Uncountable’s definition). In formulation, a typical use is predicting properties from past formulation ratio and property data and proposing the formulation to try next with Bayesian optimization or similar methods. Structuring data and defining goals and constraints are needed before the algorithm.

How much data does formulation optimization AI need?

There is no single standard. It depends on the number of variables, property variability, and whether the target is inside or outside the range of past data. Citrine’s materials argue for methods that assume little data, on the grounds that in chemistry the cost of making and testing a single sample is high and datasets tend to be small (vendor claim). For themes with little past data, vendors also present an approach that starts with DOE-like initial experiments and increases data sequentially. Start by measuring the “usable share” of past data for the target theme.

What is the difference between Bayesian optimization and design of experiments?

Design of experiments (DOE) is a method in which experiment combinations are decided in advance and run at once. Bayesian optimization is a sequential method that updates the model while looking at results and proposes, one at a time (or a few at a time), the conditions expected to yield the most if tried next. The two are not opposed, and a combination in which initial experiments are designed in a DOE-like way and then continued with Bayesian optimization is also shown in the Citrine materials mentioned above (vendor claim).

How much does formulation optimization AI cost?

The official pages of the major commercial MI platforms checked in this article’s research did not show prices. Costs are quoted individually. In terms of cost structure, comparison becomes easier if you obtain quotes split into initial costs (structuring past data, defining goals and constraints, PoC) and annual costs (license or cloud, maintenance, external support), and calculate the net benefit for each reduction rate as in this article’s model calculation.

Can we start with little past data?

Yes. However, the premise is not building a model from past data alone, but proceeding while creating data through initial experiments. Also, even with a lot of past data, records with unknown evaluation conditions or lots may be unusable. Conversely, cases have been reported where adding too much knowledge or data made the optimization worse, so it is safer to start with a small theme with a narrowed set of variables.

Can we run trials directly on AI formulation design proposals?

Avoid running trials on proposals as they are. AI does not fully understand all safety, regulatory and equipment constraints. Proposed formulations should be checked by engineers from the viewpoints of safety, regulations, equipment, cost and chemical validity, and approved before trials. Recording the reasons for approval and rejection provides material for improving the quality of the next proposals.

Summary

If you start formulation optimization AI by comparing algorithms, you risk ending up with proposals that never lead to a usable formulation. First, narrow down the target theme with four conditions (numerically measurable, costly trials, around 10 variables, stable evaluation), and define target variables and acceptable ranges, raw material upper and lower limits, the 100% total, regulated substances, the cost ceiling and equipment constraints. Align past trial data on formulation ratios, raw material lots, process conditions, evaluation conditions, results and reasons for failure, and measure the usable share. Then place an engineer’s approval between the AI proposal and the trial, and decide the days per loop and the stopping criteria. Deciding things in this order makes it clear what to buy and what you need to decide yourselves.

Published reduction rates are values for specific material systems, specific themes and specific comparison baselines. In the model calculation (placeholder figures), looking only at trial cost reductions, payback depended strongly on the reduction rate and the number of projects, and below a reduction rate of about 23% even annual costs were not covered. That is exactly why the investment decision should be made after measuring the reduction rate through retrospective verification on your own past projects and a PoC, and why shorter development periods and cost reductions should be evaluated separately. While project numbers are small, starting with data structuring and DOE is also an option.

TOMAS TECH helps plants in Thailand organize data, use AI and integrate with existing systems, tailored to their operations. You can consult us even before the product or approach has been decided, for example at the stage of narrowing down the target theme or taking inventory of past trial data. Please contact us through this form.

References

• Uncountable: Materials Informatics & the Future of Advanced Material Development
• Citrine Informatics: How to get started with AI in Paints and Coatings – and why you need to do that now (White Paper)
• BoTorch official site
• arXiv: Information Theoretic Bayesian Optimization over the Probability Simplex (arXiv:2603.09793)
• University of Cambridge Apollo: Optimization of Formulations Using Robotic Experiments Driven by Machine Learning DoE
• NIMS: MI2I (Materials Research by Information Integration Initiative)
• NIMS: About DICE
• NEDO: News release (April 13, 2020, in Japanese)
• Hitachi, Ltd.: News release (June 28, 2021, in Japanese)
• Hitachi High-Tech: Hitachi High-Tech and NOF Metal Coatings improve the efficiency and sophistication of R&D work through materials informatics (August 19, 2025, in Japanese)
• Sika AG: Investor and Media Conference – Fast Forward (November 27, 2025)
• Albert Invent: Advanced Polymer Coatings (APC) Collaborates with Albert Invent (June 23, 2026)
• Citrine Informatics official site
• GitHub: meta-pytorch/botorch
• Ax official site
• Optuna official site
• GitHub: optuna/optuna
• GitHub: SheffieldML/GPyOpt
• GitHub: scikit-optimize/scikit-optimize
• Polymerize official site
• arXiv: When Less is More: A Story of Failing Bayesian Optimization Due to Additional Expert Knowledge (arXiv:2511.16230)
• Uncountable: SCG Chemicals Accelerates Research & Development by 20% with Uncountable
• Hitachi: News release (June 30, 2025)
• Thailand BOI: A Guide to The Board of Investment 2022 Section 5
• Thailand BOI: Press Release No. 63/2564 (O.19) Enhanced Incentives to Promote R&D and HRD (June 30, 2021)