Machine learning models used for materials discovery frequently predicted physically impossible properties until now. Can Polat from Texas A&M University and colleagues have shown that accurate prediction of symmetry-forbidden responses relies on a key architectural detail: parity labels within their features. They identified a fundamental flaw impacting machine learning’s ability to accurately design new materials; without specific construction with ‘parity labels’, models often disregard essential physical symmetries.

These parity labels, information about how features reflect across space, sharply improve prediction accuracy by reducing incorrect responses by up to six orders of magnitude without affecting overall performance. Machine learning models are increasingly being used to accelerate materials discovery, replacing computationally intensive first-principles calculations, but ensuring these models respect fundamental physical laws has proven challenging. Symmetry is a vital principle governing material behaviour, specifically, how properties remain unchanged under transformations like rotation or reflection.

While much focus has been on enforcing rotational symmetry within these models, an equally key aspect often overlooked is ‘parity’, which dictates whether features change sign upon spatial inversion; consider labelling ingredients for a recipe with instructions on how they transform during cooking. Can Polat and colleagues discovered that the ability of a model to accurately predict responses forbidden by crystal symmetry depends critically on incorporating this parity information into its design via ‘parity labels’.

Parity label incorporation enhances machine learning accuracy for symmetric crystal structures

The technique central to this work carefully examined how machine learning models represent crystal symmetry through ‘parity labels’; these are tags attached to input data indicating whether a feature should change sign under certain symmetry operations, akin to labelling ingredients for a recipe with instructions on their transformation during cooking. Two thousand centrosymmetric crystals, those possessing inversion symmetry and featuring a centre point where every atom mirrors another equivalent one, were used in a systematic variation of parity label inclusion when constructing otherwise identical models, enabling direct comparison of performance. Models were paired identically except for the single design element: the inclusion of parity labels within their feature sets.

Evaluation focused on predicting the piezoelectric tensor and employed an established floating-point floor as a benchmark for accuracy. This approach bypassed reliance on training data containing explicit zeros and avoided issues associated with pre-filtering candidate crystal structures based on imperfect symmetry assessments. Scale or loss reweighting attempts failed to deliver comparable exactness; achieving this level of precision requires more than just optimisation techniques. The methodology highlights that accurate predictions depend heavily on model architecture rather than solely relying on refined parameters.

Parity Labels Enforce Physical Symmetries and Eliminate Impossible Predictions

Employing machine learning models incorporating parity labels dropped error rates for predicting physically impossible responses from between 90% and 96% to the floating-point floor, representing an improvement spanning six orders of magnitude in accuracy. Performance reached the limits of floating-point precision across two thousand centrosymmetric crystals requiring zero piezoelectric tensors, demonstrating virtually nonexistent incorrect predictions with this architectural change.

A negligible fraction of tested crystal structures received physically impossible response predictions when using models with features carrying parity labels. Symmetry constraints are often embedded within universal potentials themselves, meaning even pre-trained models carry inherent symmetries determined during their initial construction phase and can be verified via a single reflection test at model initialisation. A criterion derived from group theory, termed the ‘parity gap’, now allows proactive identification of potentially flawed architectures by assessing whether models respect these fundamental rules before training begins.

Reflection symmetry boosts machine learning performance for crystalline structures

The findings offer a pathway to building more reliable machine learning models for materials science and demonstrate that incorporating information about reflection symmetry isn’t simply optimisation but essential for physically plausible predictions. While current approaches often focus on rotational symmetries, enforcing parity, or reflection symmetry, has largely been overlooked despite its equal importance in achieving accurate results. This work centres solely on centrosymmetric crystals, those possessing inversion symmetry where every atom mirrors another equivalent one; however the implications extend beyond these specific materials. A convincing demonstration is provided of how incorporating this information dramatically improves prediction accuracy for material properties like piezoelectricity which measures a crystal’s response to mechanical stress. The research underscores the critical role of architectural design alongside parameter tuning when developing machine learning models capable of accurately predicting complex physical phenomena and opens avenues for exploring similar strategies with other types of crystalline structures.

The researchers found that machine-learned models predict physically impossible responses for some crystalline structures unless their features carry parity labels reflecting mirror symmetry. This matters because it demonstrates that including such fundamental symmetries in a model’s architecture, rather than solely through training data, ensures more reliable predictions of material behaviour.

Models incorporating these parity labels achieved accuracy at the limits of computational precision on two thousand centrosymmetric crystals where piezoelectricity must be zero, unlike those without them which frequently predicted incorrect results. The authors derived a ‘parity gap’ criterion to assess whether an architectural design respects reflection symmetry before commencing training.

👉 More information
🗞 A single design choice determines whether machine learning models of materials make physically impossible predictions
✍️ Can Polat, Mustafa Kurban, Erchin Serpedin and Hasan Kurban
🧠 ArXiv: https://arxiv.org/abs/2608.18714

More like this