Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Energy Loss Functions for Physical Systems

Authors: Oumar Kaba, Kusha Sareen, Daniel Levy, Siamak Ravanbakhsh

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our approach on molecular generation and spin ground-state prediction and report significant improvements over baselines. ... Empirical evaluation on a range of tasks, showing consistent improvement over baselines.
Researcher Affiliation Academia Sékou-Oumar Kaba , Kusha Sareen , Daniel Levy, Siamak Ravanbakhsh Mc Gill University Mila Quebec Artificial Intelligence Institute
Pseudocode No The paper describes methods and procedures in prose and mathematical formulations, but does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Code is available at https://github.com/kushasareen/energy_loss.
Open Datasets Yes First, we train diffusion models to unconditionally generate molecules in the QM9 dataset [Ramakrishnan et al., 2014]. ... We also generate large molecules with GDM and GDM-aug using the GEOM-Drugs dataset [Axelrod and Gomez-Bombarelli, 2022].
Dataset Splits Yes We construct a dataset of 10,000 training and test spin-glass Hamiltonians, each with couplings uniformly sampled from [ 1, 1].
Hardware Specification Yes Models are trained in parallel on an Nvidia Quadro RTX 8000 using the Adam optimizer. ... Runs were conducted on single 48G GPUs mainly on the Nvidia Quadro RTX 8000, A6000 and L40S. ... Training is distributed across 4 80G Nvidia A100l GPUs. ... Training takes around 5 hours on Nvidia V100 GPUs.
Software Dependencies No The paper mentions using the 'Adam optimizer' but does not specify versions for any programming languages, libraries, or other software dependencies.
Experiment Setup Yes Models are trained in parallel on an Nvidia Quadro RTX 8000 using the Adam optimizer. ... we train all models for 50 epochs. ... We conduct extensive sweeps for learning rate and positional loss weight for all losses. ... Learning rates were searched for in broadly in the range [1e-5, 1e-2] before narrowing the range to [2e-3, 4e-4]. For the positional loss weight, we choose values in [0.05, 0.1, 0.5, 0.8, 1.0, 1.5]. ... The diffusion process has 1000 diffusion steps with polynomial noise schedule and precision 1 10 5. An L2 denoising loss is used with mini-batch size 512 on GDM and 400 on EDM. We use the Adam optimizer. An EMA decay of 0.9999 is used. ... All networks are trained with a learning rate of 1 10 3 until convergence. We use the Adam optimizer with batch size 256. Temperature is set to T = 0.1 for the local energy loss.