Overview
Toshiba has developed an AI-based high-dimensional Bayesian optimization technology for automatically optimizing large numbers of parameters that would be difficult to explore manually while designing devices such as high-performance power semiconductors and advanced materials.
The design of devices and materials involves various design parameters, such as dimensions and concentrations. When designing advanced materials or high-performance devices with complex structures, performance enhancement is achieved by seeking out the optimal combination of values through the adjustment of more parameters. For example, using grid search (*1), a typical optimization method employed in manual design, if the target performance is evaluated using ten different values for each parameter, the number of combinations requiring evaluation to determine the optimal parameter values increases to 100 (10²) for two parameters and to 10 billion (10¹⁰) for ten parameters. Manually optimizing a large number of parameters is thus difficult and time-consuming. However, the developed AI enables the optimization of high-dimensional parameter vectors (*2) comprising numerous parameters, realizing automatic design based on advanced data analysis (Figure 1). This new development contributes to promote digital evolution (*3) and transformation (DE/DX) through data-driven design, contributing to the enhancement of device and material performance while improving productivity.
Toshiba applied this AI to the automatic design of a certain power semiconductor device. As a result, design parameter values were successfully found that reduced the on-resistance, which causes to power loss during operation, to two-thirds of that found using conventional search methods based on standard Bayesian optimization.
Toshiba will present the details of this technology at the 22nd International Conference on Machine Learning and Applications (ICMLA2023), which will be held in Jacksonville, Florida, USA, from December 15–17.
Development background
Conventionally, high-performance devices and advanced materials have been designed and developed based on a trial-and-error process that involves repeating (1) setting specific parameter values, (2) sample prototyping, and (3) evaluation. When multiple parameters are involved, however, the dimensionality of the parameter vector, which includes each parameter as an element, increases. This increases the number of combinations requiring evaluation to determine the optimal parameter vector values. Sample prototyping costs can be reduced by introducing a simulator that reproduces the behavior of the development and design targets, but the number of combinations is unchanged and remains large, so using simulators had not fully maximized the target performance.
In recent years, the advancement of sophisticated device functions and development efficiency has been promoted through DE/DX initiatives implementing AI and materials informatics (*4). Bayesian optimization, an AI technology, is beginning to be used to efficiently set parameter values. However, a problem with the standard approach is that it does not yield good results for large numbers of parameters.
Features of the technology
Against that background, Toshiba developed an AI-based high-dimensional Bayesian optimization technology for automated design (Figure 1) in the development of high-performance devices and advanced materials. This technology is designed to handle the numerous parameters that are challenging to address with conventional standard Bayesian optimization (Figure 2).
While standard Bayesian optimization searches for parameter values across the entire parameter space, this AI limits the search range to lower-dimensional spaces and sequentially switches among them. Furthermore, while standard Bayesian optimization uses all available data as training data for the machine learning model that predicts objective function values, this AI restricts the data to those near the low-dimensional search space. These technologies enable efficient optimization, even when there are many parameters.
This AI overcomes limitations on the number of adjustable parameters, realizes automated design based on advanced data analysis, and maximizes the design target performance. In this automated design, the manual tasks required of the designer are limited to setting an objective function based on characteristic values of the design target for which improved performance is desired and defining the search range for the parameters. This significantly enhances the efficiency of the design process and enables designers to work on a larger variety of products simultaneously and in parallel.
Applying this AI to the design of a power semiconductor device with six design parameters, Toshiba successfully automated the process of finding design parameter values that reduced on-resistance, which needs to be minimized to decrease power loss, to two-thirds of the value obtained using standard Bayesian optimization (Figure 3).
Figure 3: Results of applying this method to a power semiconductor device design problem
Future developments
In fiscal 2023, Toshiba plans to begin automated design using this AI in a power device designing department of Toshiba Electronic Devices & Storage Corporation, which handles the semiconductor and HDD businesses of Toshiba Group. The developed AI is applicable not only to the design of power devices; it can also be adapted to the automated design of various other high-performance devices and advanced materials. By accelerating DE/DX with data-driven design using this AI, Toshiba will help realize high-performance products and improved productivity in various fields.
*1: A method of setting candidate values for each parameter, evaluating performance with each combination, and finding the combination that gives the best performance.
*2: A set of numbers with each parameter as an element. Its dimensionality is, for example, 2 for two elements, and 10 for ten elements. Specifically, the vector (x1,x2,x3,x4,x5) with parameters x1,x2,x3,x4,x5 as elements is a 5-dimensional parameter vector.
*3: “Digital evolution” refers to the increased efficiency of existing businesses through digital technology.
*4: Efforts to streamline the development of materials used in products through approaches based on information science such as statistics and machine learning.