Post

Preprint! "Inverse Design of Complex Nanoparticle Heterostructures via Deep Learning on Heterogeneous Graphs" is now on ChemRxiv. In this work, we consider the problem of applying deep learning to heterogeneous nanostructures. (1/10) #ChemSky 🧪 #CompChem #physics #optics #nanoscience

Former LBNL postdoc Dr. Eric Sivonxay (@samblau.bsky.social group) led this work that began with a simple goal: designing core-shell upconverting nanoparticles (UCNPs) that can convert low-energy near-infrared light to visible and ultraviolet emissions with high brightness. (2/10)

Towards this end, we performed thousands of high-fidelity kinetic Monte Carlo simulations (using RNMC, recently published in JOSS) and compiled them into a dataset called SUNSET - Simulated Upconverting Nanoparticle Spectra for Emission Tuning. (3/10)

With SUNSET, we can learn structure-property relationships for these complex nanostructures. We considered a number of different representations for nanoparticles - simple feature matrices, voxels (for convolutional neural networks), and graphs (for graph neural networks or GNNs). (4/10)

We also designed a new heterogeneous graph representation, with one set of nodes representing UCNP dopants and another set representing interactions within a layer or between layers. We found that our hetero-GNN outperformed other models in-distribution and especially out-of-distribution. (5/10)

While testing out these different models, we realized that UCNPs and similar core-shell particles display a unique "subdivision invariance". You can define layers arbitrarily, and the particle behavior doesn't change, but the representations (including for heterographs) do change. (6/10)

To address this, we included on-the-fly data augmentation, where the same particle would be included in a training batch with multiple layers divided in multiple different ways. This data augmentation allowed us to effectively learn subdivision invariance and lower our OOD error. (7/10)

With a trained hetero-GNN, we then got to the work of optimization. Because all of the structural features were included in our graph representation, we could directly perform gradient-based optimization, rapidly identifying particles far out of distribution with exceptional properties. (8/10)