Scientific Machine Learning
Scientific Machine Learning Mini-Course
The mini-course is a short seminar series (12+ weeks starting Oct 1, 2020) with the goal of cross-pollinating ideas between the various emerging methods at the intersection of physics and machine learning.
Seminar Format: Presenters can use the opportunity to showcase a paper or two with an explicit focus on the methodology and approach. Duration: 40 minutes of methodology + 20 minutes of implementation (code) walk-through + 20 minutes of questions.
Seminar Time: Thursdays 8:30 pm to 10 pm Eastern Time
Webpage for most recent updates
Tentative Speakers and Topics
- Oct 1: Dilip Krishnamurthy, Carnegie Mellon University
Physics-Informed Neural Networks - Oct 8: Xiaowei Jia, University of Pittsburgh
Physics-Guided Machine Learning for Scientific Discovery - Oct 15: Alexander Bills, Carnegie Mellon University
Universal Ordinary Differntial Equations for Battery Performance and Reliability Predictions - Oct 22: Bharath Ramsundar, Creator of DeepChem & Stanford CS PhD
Physics-Constrained Machine Learning - Oct 29: Edwin Khoo, Institute for Infocomm Research
Physics-Aware Machine Learning for Battery Performance Predictions - Nov 5: Alok Warey, General Motors
Deep Learning for Vehicle Systems - Nov 12: Linfeng Zhang, Princeton PhD
Integrating Machine Learning with Physics-Based Modeling - Nov 19: Dhairya Gandhi, Julia Computing
Scientific Machine Learning Methods and Tools (DiffEqFlux.jl) - Nov 26: Varun Shankar, Carnegie Mellon University
Physics-Constrained Machine Learning for Fluid Flow Fields - Dec 3: Rachel Kurchin, Carnegie Mellon University
Physics-Guided Convolutional Neural Networks - Dec 10: Keith Phuthi, Carnegie Mellon University
Machine Learning Potentials and Encoding Invariances Within ML Methods
Overleaf Template for Methodology Walk-Through
Colab Notebook Format for Code Walk-Through
Funding
The seminar series is funded by the ARPA-E DIFFERENTIATE program and the Carnegie Mellon Presidential Fellowship.
Conception
In the past, I am glad to have been a panelist on the topic, Machine Learning Based Approaches to Accelerate Energy Materials Discovery and Optimization, which led to an energy focus article that summarizes recent progress, challenges, knowledge gaps, and future directions in machine learning research and applications to energy materials discovery and optimization. The core ideas from the panel discussion crystallized in my mind several aspects to carefully consider and be mindful of while applying machine learning methods to limited-data engineering challenges. Often it’s useful to synergistally stack the ML models to the extent possible with the known physics of the problem for effective learning even in low-data regimes.
In a similar vein, my recent work on Deep Learning With Physics-Based Descriptors for Electrolyte Design to Enable Ammonia Synthesis could be of interest to you. The work involves a closed-loop design methodology between computation and experiments (Manthiram group, MIT).