Santiago Miret is a deep learning researcher at Intel Labs, where he focuses on developing artificial intelligence (AI) solutions and exploring the intersection of AI and the physical sciences.
- The AI for Accelerated Materials Design (AI4Mat) workshop is scheduled for December 2nd, 2022, during the NeurIPS 2022 at the New Orleans Convention Center.
- The research focus of AI4Mat centers on the idea of Self-Driving Materials Laboratories, which promote a fully automated and closed-loop approach to designing next-generation materials design systems.
- Intel and its collaborators are contributing eight workshop papers spanning multiple relevant topics pertaining to AI4Mat.
This year at the Conference on Neural Information Processing Systems (NeurIPS), Intel Labs researchers, along with academic and corporate co-organizers from the University of Toronto, Google Research, UCSD, and Acceleration Consortium, are hosting a workshop on AI for Accelerated Materials Design (AI4Mat). The research focus of AI4Mat centers on the idea of Self-Driving Materials Laboratories, which promote a fully automated and closed approach for next-generation materials design systems.
Figure 1. A schematic representation detailing the relationship between the main components of self-driving materials laboratories. AI-Guided Design focuses on the generation of novel materials systems, while Automated Chemical Synthesis focuses on synthesizing those details in the real-world and Automated Material Characterization focuses on the effective analysis of realized material systems.
As illustrated in Figure 1, self-driving materials laboratories focus on the cross-disciplinary integration of diverse fields (materials science, machine learning, computer science, chemistry, robotics) along three primary components:
- AI-Guided Design: This component focuses on applying advanced algorithms to discover new materials and chemistries for challenging applications, such as clean energy technologies and sustainable agriculture. The primary goal is to highlight AI-based algorithmic innovations prevalent at NeurIPS, such as generative algorithms and reinforcement learning methods, that can be applied in a materials design setting.
- Automated Chemical Synthesis: This component focuses on creating automated systems that can bring new material systems into the real world through a plethora of synthesis and fabrication techniques. One particularly promising idea is fully automated synthesis systems driven by advanced robotics that can significantly increase the speed, safety, and reliability of chemical fabrication.
- Automated Material Characterization: This component focuses on automated analysis of real-world materials systems with the goal of determining their composition, structure, and functional properties. This component includes a rich intersection with computer vision, such as automated analysis of microscopic images and machine learning-based analysis of multi-modal sensors generated from different kinds of instruments.
AI4Mat’s research aims to provide the foundational technologies for Intel’s intelligent and sustainable factory of the future, as expressed by the IDM 2.0 strategy and RISE 2030 goals, especially those pertaining to sustainability. As Intel’s technology platforms continue to advance and its manufacturing footprint grows, so will the application of AI to enable state-of-the-art sustainable manufacturing. Hence, the ongoing discovery, synthesis, and characterization of diverse kinds of materials and materials-based products, such as semiconductor integrated circuits, will further highlight the potential efficiency and reliability gains of AI-infused automation platforms. Moreover, an increased manufacturing footprint demands a more sustainable technology foundation that can be amplified by the various AI capabilities, such as AI-Guided Materials Design and the discovery of sustainable chemistries that can be synthesized and analyzed at high volume to meet manufacturing and product quality needs.
In addition to spearheading the organization of the workshop, Intel researchers are also showcasing their state-of-the-art research pertaining to AI4Mat’s various themes. These contributions build on top of current research engagement with Alán Aspuru-Guzik’s Matter Lab at the University of Toronto and the recently announced engagement with MILA on responsible AI.
This year, Intel and its collaborators are contributing eight workshop papers to AI4Mat:
- MatSciML: A Flexible Framework for Scientific Machine Learning on the OpenCatalyst Dataset – This submission showcases the newly released open-source MatSciML library that aims to help machine learning researchers apply advanced deep learning to materials science design challenges on the OpenCatalyst dataset, one of the world’s largest materials datasets. In a spotlight presentation, Intel Labs researchers will provide a tutorial to the library to AI4Mat attendees on December 2nd, 2022.
- Multi-Objective GFlowNets – This submission, prepared by MILA and Intel Labs, describes a novel framework for multi-objective optimization based on GFlowNets, one of MILA’s recent algorithmic innovations for generative modeling.
- Group SELFIES: A Robust Fragment-Based Molecular String Representation – this submission, prepared by the Matter Lab and Intel Labs, showcases a novel molecular grammar that can significantly speed up the design of new molecules for drug discovery and various other impactful applications.
- Conformer Search Using SE3-Transformers and Imitation Learning – this submission, prepared by the Matter Lab and Intel Labs, applies advanced AI techniques, including geometric deep learning and reinforcement learning, to the challenge of conformer search in molecular structures. Conformer search aims to translate two-dimensional molecular formulas into three-dimensional structures that can be used to assess a molecule's properties and behavior in a variety of design use cases, such as drug discovery or functional chemistry in fabrication systems.
- Hyperparameter Optimization of Graph Neural Networks for the OpenCatalyst Dataset: A Case Study – this submission provides a framework for hyperparameter optimization on costly deep learning training workflow with different data split sizes, such as the OpenCatalyst dataset. In this submission, Intel researchers outline a method to resolve different data splits as different fidelities while optimizing the performance of the AI model as well as the computational cost required to perform the training experiment.
- On Multi-information source Constraint Active Search – this submission focuses on integrating data from different sources and qualities, such as fidelities, into a unified machine learning-based framework applied to materials design challenges.
- Human-In-The-Loop Approaches in Task Guidance – this submission introduces a novel dialog system for computer-assisted factory settings where advanced AI technologies can assist factory workers in real-world tasks in complex factory settings. The dialog systems build upon state-of-the-art natural language processing models to interact with humans while performing various tasks, thereby helping to accelerate materials synthesis and characterization.
- PhAST: Physics-Aware, Scalable, and Task-specific GNNs for accelerated catalyst design – this submission, prepared by MILA and Intel Labs, introduces a novel geometric deep learning model architecture for predicting materials properties based on data derived from expensive physics-based simulation techniques.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.