Dr. John Kitchin
Carnegie Mellon – College of Engineering
Friday, September 13, 2024
12:00 Noon
Room 120 – Meyerhoff Chemistry Building
Host: Dr. Joe Bennett
“Applications of Open Catalyst Project models in uncertainty quantification, geometry optimization and transition state search”
The Open Catalyst project provides large DFT data sets of catalytic materials and adsorbates on catalytic surfaces, a series of ML models, and pre-trained model checkpoints that can be used on their own, or as starting points for new task-specific fine-tuned models. The pre-trained models are trained on millions of DFT calculations that represent relaxation trajectories and ground-state relaxed geometries. Two outstanding questions have been how reliable are these models in predictions, and can they be used reliably for other tasks? In this talk I will present progress we have made in uncertainty quantification of these models and an application of using pre-trained models to compute reaction barriers. For uncertainty quantification, we compared several methods including ensembles, latent space and conformal methods. We find the conformal methods perform the most reliably and they can be used to help identify predictions that are out of domain and unreliable. We also used pre-trained models to calculate reaction barriers for 932 different reactions. We found a 91% success rate with the OCP model and a 28x speedup compared to just DFT. We will show results from reproducing two case studies using a pre-trained MLP that show we can find lower energy transition states in many cases through a more exhaustive search of possible pathways.