Materials Design Through Batch Bayesian Optimization with Multisource Information Fusion
- PDF / 1,481,343 Bytes
- 13 Pages / 593.972 x 792 pts Page_size
- 33 Downloads / 408 Views
https://doi.org/10.1007/s11837-020-04396-x Ó 2020 The Minerals, Metals & Materials Society
AUGMENTING PHYSICS-BASED MODELS IN ICME WITH MACHINE LEARNING AND UNCERTAINTY QUANTIFICATION
Materials Design Through Batch Bayesian Optimization with Multisource Information Fusion RICHARD COUPERTHWAITE ,1,3 ABHILASH MOLKERI,1 DANIAL KHATAMSAZ,2 ANKIT SRIVASTAVA,1 DOUGLAS ALLAIRE,2 ` YAVE1,2 and RAYMUNDO ARRO 1.—Materials Science and Engineering Department, Texas A&M University, College Station, TX 77843, USA. 2.—J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA. 3.—e-mail: [email protected]
Integrated computational materials engineering (ICME) calls for the integration of simulation tools and experiments to accelerate the development of materials. ICME approaches tend to be computationally costly, and recently, Bayesian optimization (BO) has been proposed as a way to make ICME more resource efficient. Conventional BO, however, is sequential (i.e., one-at-atime) in nature, which makes it very time-consuming when the evaluation of a materials design choice is costly. While conventional high-throughput approaches enable the synthesis and characterization (or simulation) of materials in a parallel manner, they tend to be ‘‘open loop’’ and are unable to provide recommendations of what to try next once the parallel experiment/ simulation has been carried out and analyzed. Here, we address this problem by introducing a batch BO framework that enables the exploration of the material’s design space in a parallel fashion. We augment this approach by incorporating information fusion frameworks capable of integrating multiple information sources. Demonstrating the proposed approach in the computational design of dual-phase steel, we show that batch BO can result in a significant reduction in the time and resources needed to carry out the design task. The proposed approach has wider applicability, well beyond the ICME example used to demonstrate it. Key words: Computational materials science and engineering, knowledge gradient, reification, model fusion
INTRODUCTION Integrated computational materials engineering (ICME)1 calls for the integration of various computational tools (validated against experiments) to establish quantitative process–structure–property– performance (PSPP) relationships. Inverting these relationships can accelerate the design of materials—under the key assumption that simulations are faster and cheaper than experiments. However, there are still significant challenges to this approach.
(Received June 25, 2020; accepted September 18, 2020)
Despite the assumption that simulations are cheaper than experiments, a major drawback of ICME implementations is the considerable computational cost associated with evaluating PSPP chains. This has recently been addressed through the deployment of Bayesian optimization (BO) to efficiently balance the exploration and exploitation of materials design spaces.2,3 Furthermore, most ICME frameworks tend to as
Data Loading...
