AI-Accelerated Materials Design by Abir De, Satadeep Bhattacharjee

Value of the Information & Strength of the Argument

The talks provide valuable technical insights into AI-driven materials design. Abir De’s contribution lies in proposing a learnable surrogate for combinatorial graph distances, which addresses the indexing bottleneck in large-scale molecule screening. The argumentation is rigorous, building on known computational hardness results and demonstrating empirical improvements through retrieval metrics. Satadeep Bhattacharjee’s value is in outlining a comprehensive, multi-stage pipeline that integrates generative, predictive, and interpretable models. His argumentation benefits from concrete examples, such as hydrogen storage materials, and emphasizes practical challenges like data scarcity and model economy. Both presenters argue convincingly for end-to-end differentiability and for the synergy between AI and domain-specific simulations.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, with presenters grounded in established research institutions. They mention their own publications and refer to standard concepts like graph isomorphism and subgraph isomorphism, though specific external citations are limited. The program page in the video description provides context for the broader discussion meeting, but no links to primary papers are given. The title of the video accurately describes the content, as it directly reflects the two talks on AI-accelerated materials design. No audience comments or feedback were provided for analysis.

The title accurately reflects the content: both talks are about AI-accelerated materials design, with Abir De focusing on neural graph distances for molecule retrieval and Satadeep Bhattacharjee on a generative-predictive-explaining loop for hydrogen storage materials.

8/10

The talk presents original research from two experienced researchers in AI and materials science. Methods and results are described concisely, but without external validation or peer-reviewed details. The scientific rigor appears high, but claims are not fully substantiated in the talk alone.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Chair introduces the session and the two speakers, Abir De and Satadeep Bhattacharjee.
• 01:30Abir De begins his talk on neural graph distances for molecule retrieval, framing the problem of indexing large graph databases.
• 05:00De explains the challenge of combinatorial graph matching and why traditional indexing fails, motivating end-to-end differentiability.
• 10:00Discussion of subgraph isomorphism and its NP-hard nature, and the idea of approximating distances with neural networks.
• 15:00Architecture for neural graph distance: GNN processing, permutation surrogate via doubly stochastic matrices, and distance computation.
• 20:00De presents retrieval results and a theoretical result on identically distributed entries in neural embeddings, and outlines future work.
• 22:00Chair introduces Satadeep Bhattacharjee, who begins his talk on a closed-loop materials design pipeline.
• 25:00Bhattacharjee describes the generation step using GPT-2 to propose new material compositions from hydrogen storage datasets.
• 30:00Prediction funnel: using graph neural networks (CGCNN and SRC-CGCNN) to screen candidates, reducing 1200 compositions to 118 viable ones.
• 35:00Explanation step with symbolic regression for interpretability, and mention of DFT for validation and robotic synthesis for feedback.
• 40:00Discussion of Bayesian optimization for the 'decide' step and overall integration of AI models in the materials design loop.

• Discussion Meeting: Formulating a National Research Roadmap for Accelerated Materials Design — Program page of the discussion meeting where this talk was presented, providing context and related talks.

• Discussion Meeting: Formulating a National Research Roadmap for Accelerated Materials Design — The program page lists the talk and aligns with the topics presented, reinforcing the scientific context.