Accelerated Isotopologue Reduced Partition Function Ratio Prediction with Orbital-based Deep Learning

Simon Andren, Beom Seok Kang, William Goddard, John Eiler, Anima Anandkumar

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: Quantum Chemistry, Isotope, RPFR, DFT, Quantum Mechanics

Abstract: Predictions of the reduced partition function ratios (RPFRs) of isotopologues, versions of molecules differing in position and numbers of isotopes, form a predictive framework for interpreting isotopic data from natural samples, offering insights into formation pathways and environmental conditions. However, traditional computational approaches are either computationally expensive or insufficiently accurate. Here, we employ OrbNet-Equi, a state-of-the-art orbital-based deep learning framework, speeding up predictions of RPFRs by a factor of ~1000, while maintaining accuracy comparable to density functional theory (DFT). To optimize isotopic predictions, we incorporate element-wise pooling and masking strategies. OrbNet-Equi achieves target accuracy (sub-percent for $^2$H and sub-permille for $^{13}$C, $^{15}$N, $^{18}$O) with training sets as small as 500 molecules. Using the full dataset of 100,000 molecules at 300 K, the model yields a mean absolute permille error about four times smaller than the target threshold and predicts >95% of RPFRs within the desired accuracy. Compared to other non-DFT approaches, OrbNet-Equi reduces mean absolute permille error by up to 10-fold. This establishes a computational framework capable of extending RPFR predictions to reaction networks relevant to geochemical and biochemical systems.

Submission Number: 360