Features
Pitchfest for Early Career Researchers (ECRs)
26 Jul 2023
A major challenge in developing a Deep Generative Model (DGM) for materials discovery is the high cost of evaluating the generated candidates. This typically involves conducting repetitive experiments1 or simulations in a laboratory2,
which can be labour-intensive and require a lot of resources. In a project titled "Efficient Materials Design with Fewer Domain Evaluations", Dr Li Jing, a Scientist from A*STAR’s Centre for Frontier AI Research (CFAR) addressed this challenge
by proposing two strategies.
Firstly, he selected a subset of generated candidates based on information criteria to exhibit the materials with desired properties. By focusing on this selected group, the evaluation process became more efficient and cost-effective.
Secondly, instead of adjusting all the numerous model parameters, the method would identify the most important ones and modify them. This strategy aims to reduce the model parameter space and optimise the essential parameters, hence speeding up the process of fitting desired materials (as shown in Fig. 1).
| |  Fig. 1 Generate materials with desired properties using fewer domain evaluations. | |
By combining these two strategies, the DGM could efficiently produce materials with a higher likelihood of meeting the desired criteria with less intervention from domain experts. Compared with traditional DGM training paradigms, the research outcomes
are expected to reduce experts' involvement by 50% and the framework could potentially be applied to different materials designs.
The project is accepted under Pitchfest for Early Career Researchers (ECRs)
– a platform for all researchers to connect and have potential collaborations (grant application, mentorship and ideas exchange).