Design Beyond What You Know: Material Informed Differential Generative AI (MIDGAI) for Light-Weight High-Entropy Alloys and Multi-functional Composites
This project will develop a generative AI framework for materials design that combines with discriminative models to enhance robustness and overcome data scarcity. By incorporating domain knowledge, the AI model will enable a more controllable, directed, and efficient exploration of material design spaces through advanced optimisation algorithms.
Lead PI: Ivor TSANG, A*STAR’s Centre for Frontier AI Research (CFAR)
Co-PIs:
1. TAN Teck Leong, Institute of High Performance Computing (IHPC), A*STAR
2. ZHANG Yongwei, Institute of High Performance Computing (IHPC), A*STAR
3. LIM Yee Fun, Institute of Materials Research and Engineering (IMRE), A*STAR
4. NG Chee Koon, Institute of Materials Research and Engineering (IMRE), A*STAR
5. Mohit SHARMA, Institute of Materials Research and Engineering (IMRE), A*STAR
6. Kedar HIPPALGAONKAR, Nanyang Technological University (NTU)
7. ONG Yew Soon, Nanyang Technological University (NTU)
8. GAO Huajian, Nanyang Technological University (NTU)
Collaborator:
1. ONG Shyue Ping, UC San Diego
Relying on human intuition, conventional approaches to discover advanced materials for technological breakthroughs are expensive and slow. While AI could potentially shorten discovery cycles, AI in materials science is currently limited by task-specific models and narrow data ranges, often yielding only incremental improvements.
Our project aims to develop a Materials-Informed Differential Generative AI (MIDGAI) framework to overcome these limitations and enable the discovery of new breakthroughs in (i) advanced alloys and (ii) multifunctional composites with significant weight reduction.
The MIDGAI framework combines the strengths of two types of AI models: generative models (GMs) and discriminative models (DMs). This combination allows us to enhance robustness and overcome data scarcity, to efficiently explore a wider range of materials. Unlike existing GMs that rely on and are limited by post-hoc incorporation of expert knowledge, MIDGAI incorporates domain knowledge upfront to build a consistent design space in reduced dimensions, enabling more controllable, directed, and efficient explorations of the material design spaces through advanced optimisation algorithms.
With MIDGAI, we aim to discover and prototype new light-weight materials that could significantly outperform current materials in terms of properties such as density, mechanical properties, and thermal and electrical conductivites.
Our generative AI framework can accelerate the development of novel materials by using open-source datasets and data obtained from in-house high-throughput experiments and simulations. This framework will have the potential to be generalised to different material classes and will be capable of generating novel materials that meet multiple desired properties and design criteria for various industries and applications (e.g., aerospace, transport, green buildings). This will be demonstrated on high entropy alloys and composite materials as a start, where we aim to develop light-weight prototypes with tailored properties that are suitable for dual-use applications.
