New gradient estimators cut GPU time for neural-network scientific calculations

Researchers introduced two gradient estimators designed to keep neural-network optimization trainable when gradients are weak, reporting that compact networks beat larger, fine-tuned baselines while using more than an order of magnitude less GPU time on correlated flux models. The preprint says the approach surpassed density matrix renormalization group accuracy and reached chemical accuracy for nitrogen bond breaking and iodine with explicit spin-orbit coupling. The results remain preprint findings rather than independently verified production performance, but they demonstrate a concrete route to reducing the computing needed for demanding physics and chemistry calculations.