Speaker: Sam Blau
https://chipps.lbl.gov/samuel-blau/
Title: Leveraging Novel Representations and ML Model Differentiability to Optimize Heterostructured Nanoparticles and Reaction Transition States
Abstract:
Leveraging an appropriate representation, a trained differentiable machine learning (ML) model can enable optimization with improved speed and robustness compared with non-ML approaches. In this talk, I will discuss two cases demonstrating this approach:
(1) Heterostructured lanthanide-doped upconverting nanoparticles (UCNPs) are capable of near-infrared excitation to yield visible and ultraviolet emissions, with broad applications ranging from biosensing and super-resolution microscopy to 3D printing. UCNP photophysics depends on number of layers, layer thicknesses, and dopant concentrations, defining a vast chemical design space. Despite the potential to use ML to navigate this space more efficiently, UCNPs previously had neither a viable structural representation nor sufficient data for deep learning. Here, we report efforts to overcome these challenges by combining high-throughput data generation with nanoparticle representation learning. We leverage a high-performance implementation of lanthanide energy transfer kinetic Monte Carlo with automated HPC workflows to generate the first large dataset of ~6,000 simulated spectra for UCNPs. We investigate a random forest, a MLP, a CNN, a simple GNN, and eventually converge on a physics-infused heterogeneous GNN as our best performing ML architecture. We then use the trained heteroGNN to perform gradient-based optimization of UCNP heterostructure - maximizing UV emission under 800 nm illumination as a function of number of layers and maximum nanoparticle size, identifying novel structures with far higher predicted emission than any in our training data.
(2) Computationally identifying transition states (TSes) - saddle points on the quantum potential energy surface (PES) connecting reactant and product minima - is critical to mechanistic investigations of chemical reactivity. Standard approaches optimize a discretized path of "beads" held together by spring forces, which becomes increasingly unreliable as reactions become more complex. While TS optimization is traditionally performed atop a density functional theory (DFT) PES, machine learned interatomic potentials (MLIPs) are rapidly improving and poised to replace DFT for standard atomistic simulations in the near future. Here, we report efforts to develop a novel path optimization approach which leverages MLIP speed and differentiability for improved robustness. We use a neural network to continuously parameterize the path from reactant to product, avoiding discretization. We then construct a "path loss function" calculated via a numerical integral which adaptively samples bumpier regions of the PES along the path more densely. Automatically differentiating that loss with respect to the weights of the path neural network accesses second-order PES information and enables robust iterative optimization of the continuous path via deep learning training machinery. We envision this as an important tool for mechanistic discovery and high-throughput reaction network construction in the near future.
Bio:
Dr. Samuel M. Blau is a Research Scientist at Berkeley Lab working at the intersection of computational chemistry, materials science, high-performance computing, and machine learning. He received his B.S. in 2012 from Haverford College and his Ph.D. in Chemical Physics from Harvard University in 2017. Sam has pioneered the use of self-correcting molecular simulation workflows to enable the construction of chemical reaction networks describing complex reaction cascades, e.g. those responsible for battery interphase formation and photoresist patterning. Sam's research group also develops novel datasets, representations, and models for machine learning of chemistry and materials as well as methods that leverage ML model speed and differentiability for accelerated scientific discovery.
Date and Time: May 1, 2025, 11am-noon ET / 10-11am CT
Location: Coordinated Science Lab Auditorium (B02), 1308 W. Main St., Urbana IL 61801
This will be a wonderful opportunity to connect with other experts and engage in fruitful discussions.
https://go.illinois.edu/IIDAISeminar