Material Production Condition Selection Using ML Acquisition Functions

Solution Overview

Problem

Material development methods using materials informatics are inefficient due to the large number of experimental steps required to find optimal production conditions, and existing methods using linear addition of predicted values and prediction errors result in even higher inefficiencies.

Innovation Solution

A production condition selection method utilizing a machine learning model to calculate predicted property values and errors, and an acquisition function to efficiently select production conditions for improving material properties, involving an input process, search range setting, calculation, selection, acquisition, and determination processes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a huge number of experiments are conducted to comprehensively search for optimal production conditions, then the accuracy of finding optimal production conditions is improved, but the development cost and time increase significantly

Engineering Contradiction:

Improveaccuracy of finding optimal production conditionsVSAvoiddevelopment time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-calculating predicted property values and prediction errors using a machine learning model before conducting experiments. The acquisition function is pre-designed to guide the selection of experimental points, allowing the system to start with informed predictions rather than random or grid-based experimentation. This preliminary computational preparation significantly reduces the number of actual experiments needed.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback through an iterative process where experimental results are fed back into the machine learning model to update predictions. The system continuously refines the predicted property values and prediction errors based on actual measurement data, improving the accuracy of subsequent experimental point selections. This closed-loop feedback mechanism enables the system to converge to optimal production conditions with fewer experiments.

Inventive Principle:

Principle #23Feedback

2Measurement precision

If a huge number of experiments are conducted to comprehensively search for optimal production conditions, then the accuracy of finding optimal production conditions is improved, but the development cost increases significantly

Engineering Contradiction:

Improveaccuracy of finding optimal production conditionsVSAvoiddevelopment cost

Core Design Contradiction:

Measurement precisionVSLoss of energy

Solution Approach 1:

The patent applies preliminary action by pre-calculating predicted property values and prediction errors using a machine learning model before conducting experiments. The acquisition function is pre-designed to guide the selection of experimental points, allowing the system to start with informed predictions rather than random or grid-based experimentation. This preliminary computational preparation significantly reduces the number of actual experiments needed.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback through an iterative process where experimental results are fed back into the machine learning model to update predictions. The system continuously refines the predicted property values and prediction errors based on actual measurement data, improving the accuracy of subsequent experimental point selections. This closed-loop feedback mechanism enables the system to converge to optimal production conditions with fewer experiments.

Inventive Principle:

Principle #23Feedback

3Ease of operation

If linear addition of predicted value and prediction error is used as acquisition function, then the method is simple to implement, but the number of experimental steps becomes large and efficiency is low

Engineering Contradiction:

Improveease of implementing acquisition functionVSAvoidefficiency of material development

Core Design Contradiction:

Ease of operationVSProductivity

Solution Approach 1:

The patent applies parameter changes by modifying the acquisition function to use the predicted property value divided by the prediction error (y_av / y_σ) instead of linear addition (y_av + 5y_σ). This parameter transformation fundamentally changes how the acquisition function balances exploration and exploitation, leading to more efficient selection of experimental points that target regions of high predicted performance with acceptable uncertainty.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentEP4682894A1Manufacturing condition selection method, material manufacturing method, manufacturing condition selection device, and program

Publication Date: 2026.01.21 JFE STEEL CORP

Figure 1
Figure 2
Figure 3

AI summary

A production condition selection method including: an input process (S11) of inputting an initial data group including initial production condition data and property value data regarding a material to be produced; a search range setting process (S12) of setting a search range for a production condition of the material based on the initial data group; a calculation process (S13) of calculating a predicted property value that is a predicted value of a property value of the material in the search range based on the initial data group, and calculating a prediction error that is an error of the predicted property value, using a machine learning model; and a selection process (S14) of selecting a production condition of the material to be an experimental point based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted property value, and the prediction error.