Soft Sensor AI Implementation: Quality Prediction, Data Design and RFP for Thai Plants

Soft Sensor AI Implementation: Filling the “Invisible Time” Between Lab Analyses

More and more process plants are considering an AI-based soft sensor (also called an inferential sensor or virtual sensor). Quality values such as how easily a resin flows, viscosity, concentration, moisture and reaction conversion are unknown until the lab analyzes them, and the quality of product that flows out while you wait for the result cannot be seen. During grade changes, too, the decision on whether the product is back in spec waits for lab results. Online analyzers carry large equipment costs and maintenance effort. That is when the Japanese head office asks, “Can’t AI do something about this?”

This article makes one point. Implementing a soft sensor is not about “building an AI model.” It is about deciding first which quality value to target, which lab value to treat as the ground truth, which timestamp of process data to link it to, and who fixes the model, and how, when it degrades. The first goal is not to reduce lab analysis but to fill the invisible time between lab analyses. Because models degrade over time, bias correction and retraining are built into the design from the start. Reducing lab analysis frequency is an optional final stage, taken only after parallel operation has confirmed accuracy and both quality assurance and the customer have approved the change. And the model cannot be used outside the range it was trained on (new grades, new raw materials).

First, let us separate this from similar terms. Equipment anomaly detection, which catches early signs of equipment failure, looks at “the condition of the machine.” Drift detection on SPC control charts finds process changes from the sequence of values that have already been measured. Formulation optimization AI is a development-side tool that proposes the next formulation to try. The soft sensor covered in this article is a mechanism that estimates quality values that have not yet been measured (that can only be measured in the lab) from continuously recorded data such as temperature, pressure and flow, and it differs from the three above in both purpose and inputs.

This article is written for plant managers, production managers, and quality assurance, production engineering and instrumentation staff at process plants in Thailand: chemicals, resin compounding, downstream petrochemicals, rubber, food, paper and similar. It is not a recommendation or ranking of any specific product. The cost estimate in the second half uses hypothetical figures set for this article.

What Is a Soft Sensor: Definition and How It Works

In a product introduction in Journal of the Society of Instrument and Control Engineers (August 2023 issue), Fuji Electric defines a soft sensor as “a technology that estimates values that are difficult to measure in real time, using easily collected data such as temperature, pressure and flow rate,” and writes that it is used in chemical, steel, pharmaceutical and other plants. The background to adoption is that quality-related values such as component concentration and strength take time to analyze, and if product that falls below the standard moves on to downstream processes, raw materials and energy are wasted on rework.

Toagosei’s research annual report (TREND 2022) explains it more structurally. A soft sensor is a sensor that predicts variables that are hard to measure in real time, such as reaction conversion, from process variables that can be measured in real time, such as temperature, and it has no physical parts such as a detector. The relationship between explanatory variables X (temperature, pressure, etc.) and the target variable y (reaction conversion, etc.) is expressed as a function y = f(X) and built by regression analysis of past measured data. The report also notes that because the output is only a predicted value, you need to confirm that the error against measured values is small enough.

A similar term is “virtual metrology (VM).” Dreyfus et al. (2022, International Journal of Production Research) define VM as estimating product quality directly from manufacturing process data, without physically measuring it. In a systematic review of 199 papers, they show that it has applications in manufacturing beyond semiconductors and that the components of the framework include a quality estimator and a drift detector. The name differs by industry, but the idea of “estimating quality from process data without measuring it” is common.

There are uses beyond quality as well. A 2014 paper by ABB engineers presents two cases of inferential sensors used for emissions monitoring (PEMS), and states that at a European refinery, used as a backup for the continuous emissions monitoring system, they raised the availability of the monitoring system to a level well above 97.5%. This concerns emissions estimation rather than quality, and the figure applies only to this case.

As the figure shows, for an extruder, continuous data such as barrel temperature, melt pressure, screw speed, motor current and feed rate serve as inputs, and the model estimates the quality value every few minutes (every minute in the figure). Lab analysis continues as before; its results are compared with the estimates, bias is corrected when a gap appears, and the model is retrained when degradation becomes visible. The premise of this mechanism is that the lab value remains the “ground truth.”

Why Thai Process Plants Are Considering It Now

The first reason is support from the Thailand Board of Investment (BOI). Under the “Smart and Sustainable Industry” measure in the BOI’s Investment Promotion Guide (2023 edition), “adoption of digital technology” is listed as a form of efficiency improvement for existing businesses, and one of its targets is explicitly stated as “use of AI, machine learning, big data and data analytics.” Investment in AI and machine learning can be counted toward the efficiency-improvement investment amount, and the incentive is a 3-year corporate income tax exemption capped at 50% of the efficiency-improvement investment. For machinery and automation upgrades that meet conditions such as collaboration with Thailand’s automation industry, the cap rises to 100%. A BOI briefing document dated August 4, 2026 also confirms that the same measure continues. However, whether AI-related investment is counted at its full amount or at half cannot be determined from public documents. Check eligibility and applicability with the BOI on a case-by-case basis.

According to The Nation (August 13, 2026), applications under this measure reached 1,397 from its introduction in 2023 through the first half of 2026, with total investment of more than 146 billion baht.

The second reason is the investment plans of Japanese companies. In a business sentiment survey of 504 Japanese companies in Thailand (first half of 2026), reported by MGR Online (August 17, 2026) and Pattaya Mail (August 18, 2026), of the 432 companies that answered the question on investment plans, 59% (256 companies) plan to invest in replacing existing machinery, 31% (132 companies) in improving production efficiency, and 22% (96 companies) in digital transformation. The same reports, citing BOI figures, state that Japanese-capital companies submitted more than 400 applications worth more than 55 billion baht under this measure from 2023 through the first half of 2026.

The third reason is the tough production environment. According to The Nation (July 27, 2026), Thailand’s Manufacturing Production Index (MPI) for June 2026 was 94.99, down 3.10% year on year, and capacity utilization was 57.61%. By sector, basic chemicals rose 22.35% (on demand for caustic soda and chlorine), while petroleum refining fell 7.97%. Note, however, that this was the release before the revision. The Office of Industrial Economics (OIE) revised its MPI calculation method on September 30, 2026 (The Nation, October 8, 2026), so these figures cannot be compared side by side with the revised series. With utilization below 60%, reducing off-spec product and grade-change losses within a limited production volume becomes more meaningful.

Thai petrochemical major PTT Global Chemical (GC) has also presented a “Smart Plant” and plant-wide optimization using AI and digital plant models in its news from January and July 2026 (company announcements; they do not mention soft sensors or quality estimation).

Types of Methods: PLS, Neural Networks, JIT and More

The review by Kadlec, Gabrys and Strandt (2009, Computers and Chemical Engineering) lists the methods commonly used in data-driven soft sensors as principal component analysis (PCA) combined with regression, PLS (partial least squares regression), artificial neural networks, neuro-fuzzy systems and support vector machines. The core use is predicting process variables that are available only at low frequency or through offline analysis.

PLS is widely used because of collinearity in process data. Two adjacent thermometers show almost the same value. For such data, the review states that transforming it into a low-dimensional space with less correlation using PCA or PLS is the most common approach in the process industries. In the Toagosei case, PLS was also chosen as a method “with advantages such as being able to avoid multicollinearity.”

Lawrence et al. (2024, published on arXiv as a preprint that has not been peer reviewed) divide soft sensors into knowledge-driven (white box), data-driven (black box) and hybrid (gray box). In their survey of papers from 2015 to 2023, artificial neural networks attracted the most attention among machine learning methods, and PLS was often chosen for local modeling in adaptive models. On the other hand, the methods reported to have reached industrial implementation are limited to PLS, multilayer perceptrons, wavelet NNs, SVM, RVM, Gaussian process regression and regression trees. In other words, there is a gap between the many methods proposed in research and the number of methods used in the field.

Another family is the JIT (Just-In-Time) model, which builds a local model from nearby data each time a prediction is made. This is closely related to countermeasures against degradation, so we return to it in Chapter 11.

There are also hybrids of physical models and AI. Kubosawa et al. (2022, a preprint that has not been peer reviewed) state that statistical soft sensors are weak in situations not found in past data, and propose a method that estimates composition, viscosity and other values using a dynamic simulator and reinforcement learning.

The practical conclusion is that accuracy is determined by data and operation rather than by the choice of method. With any method, accuracy will not come if you train on data with broken time alignment, and the sensor will fall out of use if degradation is not addressed. It is enough to compare methods after the data is in order.

Quality Values That Suit Soft Sensors and Those That Do Not

Kimito Funatsu (University of Tokyo, 2016 lecture material) lists the problems in soft sensor development along the flow from data collection to operation: data reliability, outliers and noise, choice of regression method, overfitting, nonlinearity, variable selection, process dynamics, model interpretation, validation, applicability domain and prediction accuracy, model degradation and maintenance, and detection of abnormal data. Turning this around shows which targets suit soft sensors and which do not.

If you push ahead when many items fall in the “Not suitable” column, you end up estimating outside the training range. Lawrence et al. write that the validity of a data-driven model may not hold when conditions deviate greatly from those it was trained on, and Toagosei, too, states that because its training data was concentrated near the target conversion, predictions in the early stage of the reaction were not reliable, and that the range in which the soft sensor is valid should be set appropriately. Using it to start up new grades or new raw materials is fundamentally the wrong application.

Three Things to Decide First: Target Quality Value, Ground-Truth Lab Value, and How It Will Be Used

Before getting into the model, write down the following three points and agree on them among the people involved.

1. The target quality value. As an example, take the MFI (melt flow index, a measure of how easily a resin flows) of an extrusion line. Narrow it to one quality value and decide which line and which group of grades it will be used for. If you extend it to multiple quality values and all lines from the start, the effort for time alignment and data checking balloons all at once.

2. The lab value used as ground truth. Even for the same MFI, values can shift depending on measurement conditions and the person in charge. Which analysis method and which record will be treated as the ground truth? How is the time of analysis kept (the sample collection time, or the time the result came out)? The quality of the ground truth sets the upper limit of model accuracy.

3. How it will be used. Will the estimate only be displayed on the operator screen, used for quality judgments, or used for control? We recommend starting with “display only.” Using it for judgments requires revising quality assurance procedures, and using it for control requires further fail-safe design. The estimate in this article also assumes “display only.”

Data Inventory: Process Data, Lab Data and Operating Records

Next, take stock of the data you have. Three types are needed: process data recorded by the DCS and PLCs, lab analysis results, and operating records such as grade changes, shutdowns, raw material lots and maintenance work.

Kadlec et al. list measurement noise, missing values, outliers, collinearity and differing sampling rates as common problems in the data, and note that dealing with them requires a great deal of manual work. The main causes of missing data are sensor failure, maintenance and removal. Outliers are checked with the 3σ method, the Hampel identifier, PCA with T², and similar methods.

The same review also points out that plant information management systems often record a value only when its change exceeds a threshold, and that if the threshold is too high, important changes are missed. Historian compression settings also affect soft sensor accuracy.

How to proceed with the inventory is covered in detail in manufacturing AI data readiness, building the data collection mechanism in our article on manufacturing data collection, and choosing a historian in our article on industrial historian selection. At plants where lab values are split between Excel and paper, considering this together with a chemical production management system that links formulations and lots with quality data avoids doing the work twice.

Also check the condition of the instruments used as inputs. Toagosei cites as a practical operating issue that correct predictions cannot be obtained when the hard sensors serving as explanatory variables degrade or malfunction. Link the calibration records and measurement recording methods of the input instruments to the soft sensor’s input tag list, referring to our articles on automatic instrument data recording and calibration management systems.

Time Alignment Determines Accuracy

What affects accuracy most is time alignment. Kadlec et al. write that because material has residence time in reactors, distillation columns and so on, it is not reasonable to link two measurements taken at the same time in different places. Delays must be compensated and the data synchronized, and that requires deep knowledge of the process. They also point out that lab quality values differ greatly in frequency from automatically measured data (multi-rate), and that in batch processes, operating time differs from batch to batch, so processing to align lengths is needed.

As a practical example, Toagosei extracted the data at the moments when the reaction liquid was sampled from time-series data saved every minute, compiled it into a table, and combined it with the reaction conversion analysis values to create training data. What matters is the “time of sampling.” If you link data by the time the result came out or the time it was entered at the lab, you end up training on misaligned input data.

The lab analysis delay and frequency assumed differ from source to source, because each describes a different site. ABB’s 2014 paper states that quality analysis is usually performed 1 to 4 times a day, with sample collection and processing causing delays of more than an hour. Lawrence et al. (a preprint that has not been peer reviewed) write that quality variable samples are taken once per shift (8 hours) or once every 24 hours, and cite the shortage of labeled data as a main issue. A digitalization lead at BASF’s Geismar plant in the United States said in LSU news (2022) that instead of waiting 12 hours for a lab sample, they want to predict it with data and AI. NTT DOCOMO BUSINESS (2025, reported by Mynavi News) cites as an issue that manual analysis can be done only about once a day. All of these describe the sites in each source and are not industry-standard values. Check your own frequency and delay against your own records.

As a prerequisite for time alignment, the clocks of the DCS, PLCs, historian and lab PCs need to be aligned. If the clocks are off, no amount of careful residence-time compensation means anything. Choosing a method and acceptance testing are summarized in our article on factory time synchronization. At the same time, agree with the lab on a practice of recording the sample collection time to the second or minute (on sample labels or as a field on the entry screen).

Model Building and Offline Validation

Once the data is in place, build a model with past data and validate it with data not used for training. Which evaluation metrics should you use? Kadlec et al. state that mean squared error (MSE) is the most common for numerical evaluation, and point out that visual evaluation, such as plots of predicted versus measured values, depends on the developer’s subjective judgment. Lawrence et al. state that soft sensor performance is commonly measured by the correlation coefficient and root mean squared error (RMSE). Ren et al. (2019) use RMSE, MAE, maximum error and relative error, and Toagosei looks at mean error and maximum error using the leave-one-out method. Looking at the maximum error, not just the average, is especially important in practice, because in quality judgments, the occasional large miss is what causes trouble.

Here is the Toagosei case, with its scope stated. The target is the company’s acrylic ester synthesis process (batch), in which the reaction must be stopped at a target conversion of 77.8 to 80.0%. Previously, analysis was timed using actual reaction times as a guide, and sometimes 2 to 3 analyses were needed per batch. Data from 32 past batches was used; the explanatory variables were reaction time and instrument values acquired automatically every minute, the target variable was the analyzed reaction conversion, the method was PLS, and evaluation used the leave-one-out method. The goal was whether the conversion could be predicted with an error of 1.1% or less against a target conversion of 78.9 ±1.1%. According to the report, predicting from reaction time alone did not keep the maximum error within 1.1%, while the soft sensor achieved an error of 1.1% or less within the target conversion range. However, because the training data was concentrated near the target, the report states that predictions in the early stage of the reaction are not reliable.

In another study, Ren et al. (2019, Sensors) tested locally weighted PLS combined with particle swarm optimization on 188 industrial production samples of gray cast iron (31 of them for testing). On the same data, the proposed method achieved an RMSE of 6.8, compared with 13.6 for global PLS. This is a relative comparison on a single cast iron dataset, and the same degree of improvement will not necessarily appear in other processes.

In offline validation, record whether the training and validation periods were split while preserving time order, the accuracy for each grade, and the operating situations with large misses, and use these as material for deciding whether to move on to parallel operation.

From Parallel Operation (Shadow Run) to Operator Display

Even if accuracy is good offline, do not show the estimates to the floor right away. First, set aside a shadow-run period: run the model on actual data, but only record the estimates without showing them to the floor, and compare them with lab values. Real-time data has more missing values and delays than historical data, and problems that were invisible offline appear here.

If the shadow run shows no problems, display the estimates on the operator screen. Even at this point, do not change the lab analysis frequency. Step 4 in the figure, “lab kept as is,” is this stage.

How the estimates are presented matters for acceptance on the floor. Fuji Electric writes that data-driven methods have a black-box aspect that industry tends to dislike, so it built screens that are easy to accept rather than simply outputting calculation results (vendor-contributed article). Lawrence et al. also cite the interpretability of purely data-driven methods as a common issue. In practice, show lab values as points over the trend of estimates so that the size of the gap is clear at a glance. Showing the main inputs driving the estimate (temperature and pressure) alongside it helps operators understand it in connection with their own operation. Also prepare a display that says “do not trust this right now” when the process is outside the training range, for example by graying out the estimate.

Designing Operation for Degradation and Retraining

Design a soft sensor on the assumption that it is at its best when built and degrades afterward. Kadlec et al. write that after deployment, regular maintenance and tuning are needed, and that because performance falls due to drift and other causes, this is compensated for by adapting or rebuilding the model. Lawrence et al. also state that the performance of online soft sensors can degrade over time, and cite as countermeasures retraining with recent data and redeploying, and bias updates that bring the output closer to lab values. Funatsu cites catalyst deterioration and changes in raw materials as causes of degradation.

There are two types of degradation, and they are handled differently. According to Kadlec et al., process drift is a change on the process side, such as pump wear, weather, raw material purity or catalyst deactivation, and its cause should be identified and removed. Sensor drift is a change on the measuring instrument side, so it is handled by recalibrating the instrument or adapting the soft sensor, without touching the process. Distinguishing the two is difficult and requires expertise.

As an example of implementing bias correction, ABB presents the following method. When a lab value is obtained, the difference from the prediction is processed statistically (suppressing the influence of noise and outliers and adjusting so that no sudden changes appear), and a correction factor is added to the output only when the difference exceeds a set threshold. Model updates, it says, should be made only when a statistically meaningful difference appears. This is because correcting every time a single lab value deviates would carry the lab-side variation straight into the model.

There are also different kinds of adaptive models. Kadlec, Grbić and Gabrys (2011) divide adaptation methods into moving-window methods, recursive adaptation methods and ensemble methods. Ren et al. (2019) explain that global models degrade over time and frequent maintenance is difficult, so they use JIT, which builds a local model for each prediction and then discards it. Funatsu also proposes “database monitoring,” in which new data is added to the database if it is useful as information and discarded otherwise. However, Funatsu’s conclusion is that “there is no universal adaptive model,” and that a model matching the type of degradation should be chosen. Note that in the survey by Lawrence et al. (covering papers from 2015 to 2023, not the adoption rate in plants), adaptive models accounted for less than one third.

The points to decide in the operational design are as follows. Who will review the gap against lab values every week? Above which threshold will bias be corrected? Under what conditions will the model be retrained? Who approves and swaps in a retrained model? What is the procedure for reverting to the previous model? Mechanisms for monitoring and swapping models are covered in detail in our guide to factory AI model monitoring.

Reducing Lab Analysis Frequency Comes Last

“Being able to reduce lab analysis” is the soft sensor benefit people are most likely to expect. However, it should not be the first goal.

The Toagosei paper writes that it is “considered fully possible to reduce analysis to a single analysis for confirmation.” This is an expectation stated in the paper, not a report of having actually reduced the number of analyses in operation. This distinction matters.

Changing the frequency requires at least the following stages. Record the difference from lab values for a sufficient period through parallel operation. Have the quality assurance department define the judgment procedure and how records are handled. Explain the change to the customer and obtain approval. In particular, if inspection certificates carry actual lab measurements, whether estimates can replace them and whether the analysis frequency can be changed depend on the agreement with the customer. Some customers cannot allow a change in frequency because of their certificate requirements. For the flow of issuing and approving certificates, see our article on inspection certificate issuance control.

In the pharmaceutical field, there is the PAT (Process Analytical Technology) guidance issued by the US FDA in October 2004, which encourages companies to voluntarily pursue innovative manufacturing and quality assurance. We have not been able to confirm whether the body of this guidance mentions soft sensors. Funatsu’s lecture material (2016) explains that soft sensors make real-time release testing (RTRT) possible and that final product testing could become unnecessary, but this is a pharmaceutical topic in a researcher’s lecture material and does not apply as is to chemical or resin plants.

How to Read Products and Cases [Vendor Claims]

Below we introduce products and cases. All are announcements by manufacturers or adopting companies, or press reports based on them, and none have been verified by a third party. For none of the products could we confirm a price in public information (not disclosed). The following is not a recommendation or ranking.

As a supplementary note, Fuji Electric states that its tool can greatly improve workload and working time (by 70% or more), but it does not state what this was compared against or the measurement conditions. Seeq’s External ML is described as a separate license.

Now for the cases. According to Hydrocarbon Processing (July 20, 2026, an article based on a press release), Hyperion Systems Engineering put into operation soft sensors that estimate melt flow rate (MFR) and xylene solubles in real time at HELLENiQ ENERGY’s polypropylene plant in Thessaloniki, Greece. The sensors link with the DCS, a real-time database, LIMS and other systems, and output estimates every few minutes without installing new analyzers. According to HELLENiQ ENERGY, off-spec product fell by 20%, grade change time was shortened by 25%, and this led to an estimated annual profit improvement of €350,000 and a reduction in CO₂ of 10.5 kt per year. These are figures announced by the vendor and the adopting company for a single plant (a PP plant in Greece), and they should not be used as an industry benchmark or as your own expected results.

According to Industria Química (October 5, 2026), BASF runs a machine learning model that estimates quality parameters in real time from process data in a distillation process at its Antwerp site in Belgium, as an edge app on Siemens Industrial Edge. The benefits cited are operation closer to limits, reduced analysis time and sampling costs, and data provision to APC, but no figures on effects were reported.

As a Japanese example, NTT Communications (as it was then known) and Mitsui Chemicals used deep learning to estimate “gas X concentration” 20 minutes ahead in the manufacture of a gas product, from 51 types of process data including raw material temperature, pressure and flow. The accuracy was reported as an average error within 3%FS (reported by Weekly ASCII, September 2016). In addition, Shana (February 8, 2025) reported the unveiling of an AI virtual sensor at Iran’s Nouri Petrochemical Complex, but with no figures on effects.

There are three yardsticks for reading these. Which quality value, in which process? Is the comparison baseline (what it was compared with before adoption) stated? Is the mechanism for degradation and retraining explained?

Requirements to Include in the RFP

Before asking vendors for quotes, write the following items into the RFP. If you request quotes while these remain vague, each vendor’s quote will rest on different assumptions and they cannot be compared.

The execution environment is an item often decided by whether data can leave the control network. For how to think about running on the edge, see our edge AI factory deployment guide.

PoC and Acceptance Criteria

Structure the PoC in two stages. The first stage is offline validation with historical data: build a model with time-aligned data and check accuracy on a period not used for training. The second stage is parallel operation for a set period: record estimates with real-time data and compare them with lab values. Even if a model passes offline, missing data and delays can affect it in parallel operation, so the second stage is not skipped.

There are no industry-defined numerical acceptance criteria. Your company sets them. One approach, for example, is to set “the proportion of cases in which the difference from the lab value falls within a certain fraction of the quality specification width” and evaluate it for each grade and each operating state (steady state, during grade change). Record the maximum error and the situations with misses, not just the average. How to read PoC costs and how to decide whether to move to production are summarized in our article on AI PoC costs and success criteria.

Also check in the PoC how the estimates and on-screen warnings behave right after a grade change, right after a raw material lot change, when an input instrument has missing data, and under operating conditions that were not in the training period.

Model Cost and Benefit Estimate (Hypothetical Figures)

The figures from here on are all hypothetical figures set for this article. We could not confirm official prices for soft sensors in public information, so the costs are also placeholders. Replace them with your own figures.

Assumed plant: A Japanese-affiliated resin compounding plant in Thailand. Three extrusion lines, operating 24 hours a day, 330 days a year. The quality value is MFI. Lab analysis is done every 2 hours on each line (12 times a day), with about 40 minutes from sample collection to result (a placeholder for this article). The labor cost is 300 THB per hour.

Current state (baseline amounts, per year):

Check: for item 2, 200 × 500 = 100,000 kg, × 40 = 4,000,000 THB. For item 3, 11,880 × 150 = 1,782,000 THB.

Configuration A (MFI soft sensors on 3 lines): Uses data from the existing DCS/PLCs and historian and displays the estimates on the operator screen. Lab analysis frequency is not changed, and the estimates are not used for control. Initial cost 2,000,000 THB, annual cost 400,000 THB (model maintenance, retraining, licenses).

Configuration A does not count any reduction in lab analysis.

Sensitivity (if accuracy falls short and the benefit is halved): The benefit is 470,000 THB, the net benefit is 470,000 − 400,000 = 70,000 THB per year, and payback is 2,000,000 ÷ 70,000 ≈ about 28.6 years, meaning it effectively never pays back. Model accuracy, and whether the estimates are actually used on the floor, decide the investment case.

Configuration B (on top of A, review lab analysis frequency): After confirming accuracy through 6 months or more of parallel operation and obtaining approval from quality assurance and the customer, lab analysis is changed from every 2 hours to every 4 hours. This is evaluated as the difference from A. Additional initial cost 500,000 THB (validation, documentation, approval procedures), additional annual cost 100,000 THB.

The benefit of Configuration B is zero without approval from quality assurance and the customer. Some customers cannot allow a change in frequency because of their inspection certificate requirements. Do not include this benefit in the investment decision before approval is obtained.

Rules to avoid double counting (assumptions of this estimate):

• (a) Item 1 (off-spec product during steady operation) and item 2 (transition losses during grade changes) are separate cost buckets. Item 1 does not include losses during grade changes.
• (b) The lab analysis reduction (item 3) is included only in Configuration B, not in Configuration A or the sensitivity case.
• (c) Reducing lab analysis does not increase the off-spec reduction benefit. Configuration B adds no further benefit to items 1 or 2.

What this estimate does not include: Adding or removing online analyzers, use for control (closed loop), energy savings, and BOI tax incentives (eligibility to be confirmed individually).

What the estimate tells us: Configuration A alone takes about 3.7 years to pay back, and does not pay back if accuracy falls short. What most affects the benefit is whether lab analysis frequency can be reviewed after accuracy has been confirmed through parallel operation (Configuration B). However, that presupposes approval from quality assurance and the customer, and it is not the first goal. We recommend this order: start with Configuration A and focus on “filling the invisible time,” then consider Configuration B once the estimates are being used on the floor and a record of accuracy has built up.

Frequently Asked Questions and Summary

What is a soft sensor?

It is a mechanism that estimates quality values that can only be measured by lab analysis (MFI, viscosity, concentration, moisture, reaction conversion, etc.) from process data that is recorded continuously, such as temperature, pressure and flow. It has no physical detector, and the relationship y = f(X) is built as a model from past measured data. Because the output is only a predicted value, it is used while checking the error against lab values.

How accurate can AI quality prediction be?

There is no general figure. Accuracy varies greatly with the process, the quality of lab values, time alignment and operation, and published accuracy figures are values for specific processes and conditions. Confirm it through offline validation and parallel operation (PoC) with your own data.

Is virtual metrology different from a soft sensor?

The idea is shared: both “estimate quality from process data without physically measuring it.” Academic reviews also show applications of virtual metrology (VM) in manufacturing beyond semiconductors.

Can lab analysis be reduced?

It is possible, but it should not be the first goal. Only after accuracy has been confirmed through parallel operation, quality assurance procedures have been defined and the customer has approved can the frequency be changed. Some customers cannot allow a change in frequency because of their inspection certificate requirements. Read the reductions shown in research as expectations, distinct from operating results.

How quickly does a quality prediction model degrade?

There is no general period. The causes and speed of degradation, such as catalyst deterioration, changes in raw materials and input instrument drift, depend on the process. Keep monitoring the gap against lab values, design the correction and retraining operation from the start, and confirm how often intervention is needed during the PoC’s parallel operation.

Can we start with limited data?

If there are few lab values to serve as ground truth, both model accuracy and the reliability of validation fall. Research also cites the shortage of labeled data as a main issue. A realistic approach is to first take stock of how well lab values, sample collection times and process data line up, and if they fall short, start by recording sample collection times.

Which should we choose, a soft sensor or an online analyzer?

Think in terms of roles, not replacement. ABB writes that inferential sensors have lower initial investment and maintenance costs than analyzers, but this is a qualitative claim in a vendor-contributed article. In some cases, installing an analyzer is difficult because of cost, the properties of the liquid to be measured, or regulations (Toagosei). On the other hand, a soft sensor cannot be used outside its training range and needs lab values or analyzer values as ground truth. For processes with many new grades, or uses that require the actual measured value itself, plan on the premise of keeping analyzers or lab analysis.

Summary

• A soft sensor is a mechanism that estimates quality values that can only be measured in the lab from continuously recorded data, and it is distinct from equipment anomaly detection, SPC and formulation optimization
• Implementation is not about “building an AI model” but about first deciding the target quality value, the lab value used as ground truth, the timestamp linkage, and how to fix the model when it degrades
• What affects accuracy most is time alignment. Recording sample collection times and synchronizing the clocks of each system are prerequisites
• Models degrade. Build bias correction, retraining, approvers and version control into the design from the start
• The model cannot be used outside its training range (new grades, new raw materials)
• Figures for products and cases are announcements by vendors and adopting companies. Check the comparison baseline and conditions
• In the hypothetical estimate, the display-only Configuration A takes about 3.7 years to pay back and does not pay back if accuracy falls short. Reviewing lab analysis frequency (Configuration B) has a large benefit, but that benefit is zero without approval from quality assurance and the customer

TOMAS TECH helps factories in Thailand with data preparation, AI adoption and integration with existing systems, tailored to their operations. You are welcome to consult us even at the stage of deciding which quality value to start with, taking stock of whether time alignment is possible with your current process data and lab values, or organizing RFP requirements. Please contact us through this form.

References

• Fuji Electric: Soft sensor “estimation model building / calculation tool” (Journal of the Society of Instrument and Control Engineers, Vol. 62, No. 8, August 2023)
• Toagosei: Managing reaction conversion with a soft sensor (TREND 2022, No. 25)
• Kadlec, Gabrys, Strandt: Data-driven Soft Sensors in the Process Industry (Computers and Chemical Engineering, 2009)
• Kadlec, Grbić, Gabrys: Review of adaptation mechanisms for data-driven soft sensors (Computers and Chemical Engineering, 2011)
• Ren, Song, Chu: An Improved Locally Weighted PLS Based on Particle Swarm Optimization for Industrial Soft Sensor Modeling (Sensors, 2019)
• arXiv: Machine learning for industrial sensing and control: A survey and practical perspective (arXiv:2401.13836, preprint)
• arXiv: Soft Sensors and Process Control using AI and Dynamic Simulation (arXiv:2208.04373, preprint)
• Funatsu: Soft Sensor: Chemoinformatic Model for Efficient Control and Operation in Chemical Plants (Strasbourg Summer School in Chemoinformatics 2016)
• Bonavita, Ciarlo (ABB): Inferential Sensors for Emission Monitoring: an Industrial Perspective (2014)
• Dreyfus et al.: Virtual metrology as an approach for product quality estimation in Industry 4.0 (International Journal of Production Research, 2022)
• FDA: PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (October 2004)
• Thailand BOI: Measure for Industrial Upgrades towards Smart and Sustainable Industry (Investment Promotion Guide 2023)
• Thailand BOI: BOI Opportunity & Incentives for the Digital Industry (August 4, 2026)
• The Nation Thailand: Thailand sees smart and green applications top 146 billion baht (August 13, 2026)
• The Nation Thailand: Refineries, motor vehicles and palm oil pull June industrial output lower (July 27, 2026)
• The Nation Thailand: Goldman Sachs sees Thai first-half 2026 growth revised to 2.9% (October 8, 2026)
• MGR Online: Japan Accelerates Production Base Upgrades in Thailand (August 17, 2026)
• Pattaya Mail: Japanese factory upgrades open new high-value opportunities for Thai suppliers (August 18, 2026)
• PTT Global Chemical: News (January 13, 2026)
• PTT Global Chemical: News (July 20, 2026)
• Mynavi News TECH+: NTT DOCOMO BUSINESS AI Soft Sensor implementation support package (November 13, 2025)
• AspenTech: Aspen Inferential Qualities
• Siemens: SITRANS Soft Sensor Engine IQ
• Yokogawa: AI Product Solutions
• Seeq Knowledge Base: Machine Learning Tools FAQ
• Hydrocarbon Processing: Hyperion deploys HYPPOS AI soft sensors at HELLENiQ ENERGY polypropylene plant in Greece (July 20, 2026)
• Industria Química: BASF aplica inteligencia artificial al control de procesos de destilación en Amberes (October 5, 2026)
• Louisiana State University: Chemical Industry Leader BASF Taps LSU to Create Soft Sensors Using AI (September 20, 2022)
• Weekly ASCII: Quality prediction by NTT Communications and Mitsui Chemicals (September 2016)
• Shana: RIPI unveils innovative virtual sensor technology (February 8, 2025)