Generative AI for Discovering Novel Materials || Semi-analytical NN || Nov 21, 2025

Value of the Information & Strength of the Argument

The talks provide valuable insights into the application of generative AI to materials discovery and the development of hybrid numerical methods. The first talk demonstrates a practical pipeline from data collection to candidate generation, with some experimental validation. The argumentation is supported by references to established databases and methods. The second talk offers a novel combination of ADM and PINN, with theoretical convergence guarantees. However, the presentations are brief and lack detailed derivations or extensive validation, limiting the depth of the argumentation.

Scientific Rigor, Source Quality, Title Accuracy

The speakers are from reputable institutions and present work that appears methodologically sound. The first talk references known databases (Materials Project, OQMD, ICSD) and uses established AI techniques (CDVAE, LLM fine-tuning). The second talk builds on well-known numerical methods (ADM, PINN). However, no specific sources are cited in the video, and the description provides only abstracts. The title accurately reflects the content, though it is broad. The seminar is part of an academic series, suggesting a level of rigor, but the lack of detailed citations limits verification.

The title accurately reflects the content, covering generative AI for materials discovery and a semi-analytical neural network approach.

7/10

The seminar presents original research with a clear methodology, but lacks detailed technical depth and external verification. The speakers are credible researchers from reputable institutions, and the content is consistent with current AI for materials science literature.

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

• 00:00Introduction to the seminar and first speaker, Dibakar Datta.
• 02:00Overview of energy storage challenges and the need for new materials.
• 05:00Explanation of battery architecture and the role of anode/cathode materials.
• 10:00Discussion on the limitations of existing batteries and the potential of 2D materials.
• 15:00Introduction to the generative AI framework: combining CDVAE and LLM.
• 20:00Details on data collection from databases and preprocessing.
• 25:00Fine-tuning of LLM with LoRA and quantization for crystal generation.
• 30:00Explanation of the Crystal Diffusion VAE and its graph representation.
• 35:00Results: generation of novel materials and DFT validation.
• 40:00Transition to second speaker, Arihant Patawari, on semi-analytical neural networks.
• 45:00Introduction to time fractional reaction-diffusion models and ADM.
• 50:00Combination of ADM and PINN, and convergence analysis.

• Materials Project — Database used for collecting crystal data for training generative models.
• OQMD — Open Quantum Materials Database used for data collection.
• ICSD — Inorganic Crystal Structure Database used for data collection.
• Meta Llama 3.1 — Base LLM model fine-tuned for crystal generation.

• Materials Project — Widely used database for computational materials science.
• Crystal Diffusion Variational Autoencoder — The CDVAE method is a known approach for generative materials design.