A material process parameter self-adaptive optimization method and system based on dynamic bayesian optimization

Abstract

The application discloses a material process parameter self-adaptive optimization method and system based on dynamic Bayesian optimization, and belongs to the field of artificial intelligence assisted material research and development. In view of the problems of high cost and easy falling into local optimization under complex phase change in traditional material research and development, the application constructs an initial set by acquiring material historical process and performance data; a nonlinear proxy model is constructed by using Gaussian process regression and a composite kernel function fusing smooth and rough priori; a function acquisition strategy is dynamically switched according to a prediction variance to balance global exploration and local utilization; within the extracted physical equipment hard constraint boundary, a multi-start quasi-Newton method is used to perform global optimization, and the next round of recommended experimental parameters are output and closed-loop iteration is supported. The application realizes accurate optimization of multi-dimensional parameters with few experimental times, and is mainly used for complex chemical and material process parameter optimization such as polymerization and alloy smelting.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and materials science, and in particular to an adaptive optimization method and system for material processing parameters based on dynamic Bayesian optimization for small sample environments. Background Technology

[0002] In physics, chemistry, and materials research and development (such as polymer synthesis and special alloy smelting), finding optimal process parameters or the best formulation is an extremely challenging task. Traditional design-of-experiment (DoE) methods, such as orthogonal experiments or response surface methodology (RSM), often encounter the "curse of dimensionality" when dealing with complex reaction systems that are multivariable, nonlinear, and strongly coupled. Researchers typically have to rely on experience for high-throughput, intensive, and blind trial-and-error, resulting in high depreciation costs for experimental materials and instruments, and extremely long development cycles.

[0003] In recent years, machine learning technology has made significant progress in various fields, but deep learning models are extremely dependent on massive amounts of training data. However, in real materials and chemical engineering experiments, due to the extremely high cost and time involved in a single experiment, research teams typically only obtain a dozen to several dozen sets of extremely valuable "small sample data." Traditional neural networks are prone to overfitting on such sparse data, failing to provide reliable generalization predictions. Furthermore, real material synthesis processes are often accompanied by physical phase transitions such as glass transition and crystallization, causing discontinuous and drastic abrupt changes in the response surface of the target yield in localized areas, making it difficult for conventional surrogate models with a single kernel function to achieve high-precision fitting. In the optimization decision-making stage, traditional Bayesian optimization algorithms, if employing a single acquisition function strategy, are prone to prematurely getting trapped in local optima or wasting extremely valuable experimental resources by overexploring unknown regions. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive optimization method and system for material process parameters based on dynamic Bayesian optimization, aiming to solve the problems of low modeling accuracy for small samples, difficulty in fitting complex phase transition characteristics, and easy getting trapped in local optima in the optimization process of existing technologies in material research and development.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An adaptive optimization method for material process parameters based on dynamic Bayesian optimization, comprising the following steps: 1. Obtaining historical experimental data containing material process parameters and their corresponding performance indicators to construct an initial dataset; 2. Based on the initial dataset, constructing a nonlinear surrogate model using a Gaussian process regression algorithm with the process parameters as input features and the performance indicators as target predicted values, and outputting the predicted mean and predicted variance; 3. Calculating the acquisition function value of each candidate data point in the parameter space based on the predicted mean and predicted variance, wherein the acquisition function is used to quantify the potential experimental benefits of the candidate data points; 4. Extracting the safe operating limits of the actual physical experimental equipment and converting them into hard constraint boundaries of the process parameter space; 5. Within the hard constraint boundaries, using the acquisition function value as the optimization objective, performing global optimization using a multi-starting-point finite-memory quasi-Newton method to obtain the optimal candidate data point that makes the acquisition function value reach a global extreme value; 6. Outputting the optimal candidate data point as the recommended process parameters for the next round of physical experiments.

[0006] Furthermore, in the step of constructing a nonlinear surrogate model using the Gaussian process regression algorithm, a composite kernel function containing smooth priors and coarse priors is adopted. The composite kernel function includes a radial basis function for fitting a smooth continuous response surface and a Matern kernel function for capturing local non-smooth abrupt changes caused by material phase transitions. The system adaptively allocates the weights of the radial basis function and the Matern kernel function according to the local gradient change rate of the historical experimental data.

[0007] Furthermore, in the step of calculating the acquisition function value, a dynamic adaptive acquisition function switching mechanism is adopted: in the early stage of optimization, if the global maximum prediction variance in the parameter space is greater than the preset uncertainty threshold, the confidence upper bound strategy is adopted as the acquisition function to perform parameter space exploration; when the global maximum prediction variance is less than the uncertainty threshold for a preset number of consecutive rounds, it is automatically switched to the expectation enhancement strategy or the probability enhancement strategy to perform high-density fine-tuning near the local optimum.

[0008] The present invention also provides an adaptive optimization system for material process parameters based on dynamic Bayesian optimization, comprising: a data management module, a model building module, and an intelligent decision-making module.

[0009] The beneficial effects of this invention are as follows: Overcoming the limitations of small sample data: Gaussian process regression achieves high-precision modeling with extremely small datasets, and accurately quantifies uncertainty through output prediction variance, effectively avoiding the overfitting problem of neural networks. Improving the fitting capability of complex phase transition systems: Introducing a composite kernel function mechanism that integrates radial basis kernels and Matern kernels ensures smooth fitting of conventional regions while accurately characterizing local abrupt changes caused by physical phase transitions. Balancing global exploration and local utilization: An innovative dynamic acquisition function switching mechanism combined with a multi-starting point anti-trapping optimization algorithm approximates the globally optimal formula with the fewest physical experiment iterations, greatly reducing trial-and-error costs. Ensuring the safety of engineering implementation: Directly mapping equipment limits to the Lagrangian hard constraint boundary of the optimization process guarantees 100% executability of AI-recommended parameters. Attached Figure Description Figure 1 This is the main flowchart of an adaptive optimization method for material process parameters based on dynamic Bayesian optimization provided in this embodiment of the invention; Figure 2 This is a flowchart of the dynamic adaptive switching logic of the acquisition function provided in the embodiment of the present invention; Figure 3 This is a system module structure diagram provided in an embodiment of the present invention. Detailed Implementation To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0010] This invention provides an adaptive optimization method for material process parameters based on dynamic Bayesian optimization, which is executed by a computer device.

[0011] Step S1: Data Initialization. Acquire historical experimental data containing material processing parameters and their corresponding performance indicators. For example, in the melt polycondensation process of thermoplastic polymers (TPU), the input features are reactor temperature and system pressure, and the performance indicator is polymer yield. The system uses methods such as Latin hypercube sampling to generate an initial small amount of baseline data, which is then entered into the system.

[0012] Step S2: Construct a surrogate model using the composite kernel function. Based on the dataset above, a surrogate model is constructed using Gaussian process regression (GPR). Considering the physical properties of polymers that may undergo drastic phase transitions such as glass transition or crystallization at specific temperatures, the system automatically assigns higher weights to the Matern kernel function, utilizing its introduced roughness parameter to accurately capture local non-smooth abrupt changes. Simultaneously, the model outputs not only the mean predicted yield but also the predicted variance representing the confidence interval.

[0013] Step S3: Dynamically calculate the collected function value. This embodiment introduces an adaptive active learning game mechanism: In the early stages of exploration, there are many cognitive blind spots in the global parameter space, and the prediction variance is extremely large. At this time, the system automatically attaches the upper confidence bound (UCB) strategy. The UCB strategy will drive the system to recommend temperature and pressure coordinates with extremely uncertain (extremely large variance) models, clearing the blind spots in the parameter space in the most efficient way. After several rounds of iteration, if the global maximum prediction variance converges to below the threshold, the system determines that the "main peak" has emerged, and automatically switches the strategy to the expected improvement (EI) or probability improvement (PI) strategy. At this time, the algorithm concentrates computing power on high-density sampling near the middle of the "main peak" until it approaches the global optimum.

[0014] Step S4: Global Optimization from Multiple Starting Points within Constraint Boundaries. The limiting parameters of the physical equipment are extracted as hard constraints, such as temperature limited to 100-200℃ and pressure limited to 1-5 atm. Within the defined multidimensional constraint space, the system invokes the finite-memory quasi-Newton method (L-BFGS-B) and randomly generates 25 orthogonally distributed initial coordinate points as optimization starting points. The system performs gradient search in parallel, compares the local extrema of convergence for each path, and extracts the coordinate points that maximize the gain of the acquisition function as recommended parameters.

[0015] Step S5: Human-Machine Collaborative Closed-Loop Iteration. Researchers input the system-recommended parameters (e.g., temperature 161.39℃, pressure 3.00 atm) into the actual reaction equipment to perform physical verification. After the experiment, the actual yield measurement results are added to the initial dataset as new data nodes, triggering the surrogate model to retrain. Through this closed-loop iteration of "prediction-experiment-feedback-re-prediction," the model becomes increasingly accurate, ultimately solving the cost problem of high-throughput trial and error.

[0016] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive optimization method for material processing parameters based on dynamic Bayesian optimization, characterized in that, Includes the following steps: An initial dataset is constructed by acquiring historical experimental data containing material processing parameters and their corresponding performance indicators. Based on the initial dataset, a nonlinear surrogate model is constructed using the Gaussian process regression algorithm, with the process parameters as input features and the performance index as the target predicted value, and the predicted mean and predicted variance are output. The acquisition function value of each candidate data point in the parameter space is calculated based on the predicted mean and predicted variance, and the acquisition function is used to quantify the potential experimental benefits of the candidate data points. The safe operating limits of actual physical experimental equipment are extracted and transformed into hard constraint boundaries of the process parameter space. Within the hard constraint boundaries, with the acquisition function value as the optimization objective, global optimization is performed using the multi-starting point finite memory quasi-Newton method to obtain the optimal candidate data points that make the acquisition function value reach the global extreme value. The optimal candidate data points will be output as recommended process parameters for the next round of physical experiments.

2. The adaptive optimization method for material process parameters based on dynamic Bayesian optimization according to claim 1, characterized in that, In the step of constructing a nonlinear surrogate model using the Gaussian process regression algorithm, a composite kernel function containing smooth priors and coarse priors is adopted. The composite kernel function includes a radial basis function for fitting a smooth continuous response surface and a Matern kernel function for capturing local non-smooth abrupt changes caused by material phase transitions. The system adaptively allocates the weights of the radial basis function and the Matern kernel function according to the local gradient change rate of the historical experimental data.

3. The adaptive optimization method for material processing parameters based on dynamic Bayesian optimization according to claim 1, characterized in that, In the step of calculating the acquisition function value of each candidate data point in the parameter space based on the predicted mean and predicted variance, a dynamic adaptive acquisition function switching mechanism is adopted: in the early stage of optimization, if the global maximum predicted variance in the parameter space is greater than a preset uncertainty threshold, a confidence upper bound strategy is adopted as the acquisition function to perform parameter space exploration; when the global maximum predicted variance is less than the uncertainty threshold for a preset number of consecutive rounds, the acquisition function is automatically switched to an expectation enhancement strategy or an increase probability strategy to perform high-density fine-tuning near the local optimum.

4. The adaptive optimization method for material processing parameters based on dynamic Bayesian optimization according to claim 1, characterized in that, The step of performing global optimization using the multi-starting-point finite-memory quasi-Newton method within the hard constraint boundary includes: randomly generating multiple orthogonally distributed initial coordinate points as optimization starting points within the multidimensional constraint space composed of the high and low limits of the process parameters; for each initial coordinate point, performing gradient search in parallel using the finite-memory quasi-Newton method, and recording the local extrema when each search path converges; comparing all converged local extrema, extracting the coordinate point with the largest gain as the optimal candidate data point to overcome the local optimum trap.

5. The adaptive optimization method for material process parameters based on dynamic Bayesian optimization according to any one of claims 1 to 4, characterized in that, It also includes a human-machine collaborative closed-loop iterative step: receiving the actual performance indicators measured after the user inputs the recommended process parameters into the real physical equipment to perform physical verification; combining the recommended process parameters and the actual performance indicators into a new data entry and appending it to the initial dataset; The surrogate model update instruction is triggered, and the hyperparameters of the Gaussian process regression surrogate model are retrained using the updated dataset to achieve incremental learning and model accuracy calibration.

6. An adaptive optimization system for material processing parameters based on dynamic Bayesian optimization, characterized in that, include: The data management module is used to acquire historical experimental data containing material process parameters and their corresponding performance indicators to construct the initial dataset; The model building module is used to construct a nonlinear surrogate model based on the initial dataset using the Gaussian process regression algorithm and output the predicted mean and predicted variance. The intelligent decision-making module is used to calculate the acquisition function value based on the predicted mean and predicted variance, and to perform global optimization using a multi-starting point finite memory quasi-Newton method within the hard constraint boundary set based on the physical equipment limits, so as to obtain the optimal candidate data point as the recommended process parameters for the next round of physical experiments.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images