An AI system learns to model chemical bonds breaking across many molecules

A model trained on 22,000 molecular structures was fine-tuned to describe unfamiliar molecules with near multireference accuracy at lower repeated computational cost.

Editorial illustration — not from the study.

Describing a chemical bond as it breaks is difficult because the molecule’s electronic structure can require methods that account for several competing arrangements at once. Those methods are computationally expensive and normally repeat much of the calculation for every molecule.

In a Nature Communications study, researchers introduced Orbformer, a model pretrained on 22,000 equilibrium and dissociating molecular structures. They report that fine-tuning it on previously unseen molecules produced an accuracy–cost balance comparable to traditional methods for these difficult electronic structures, including bond dissociations and Diels–Alder reactions.

How the model handles breaking bonds

Orbformer was pretrained on 22,000 molecular structures, covering both stable molecules and molecules undergoing bond dissociation. When fine-tuned on unfamiliar molecules, it reached an accuracy–cost ratio comparable to classical multireference methods. On established benchmarks, more challenging bond-breaking tests and Diels–Alder reactions, the researchers report that Orbformer was the only method tested to consistently reach chemical accuracy, defined here as 1 kilocalorie per mole.

Why reusable calculations matter

The result supports a different way to use high-cost quantum chemistry: a model can learn recurring features of electronic structure across molecules, rather than paying the full computational cost independently for every system. If the reported performance holds across the tested types of chemistry, this could make accurate calculations for difficult bond-breaking processes more reusable and efficient.

Tests and remaining questions

This is a computational study reported in a Nature Communications research article. The evidence comes from pretraining, fine-tuning and comparisons on established benchmarks, bond-dissociation problems and Diels–Alder reactions. The abstract does not give the detailed errors, computational costs, training procedure or range of molecules tested, so it does not establish how the approach will perform outside these evaluations or how much cost it saves in specific use cases.

A portable light-based device screens alcohol and water in commercial samples

A compact near-infrared device uses lead-free perovskite materials to screen alcohol and water content in commercial samples. Its circuit turns optical signals into five visual levels, covering alcohol concentrations of 30–60% and water contents of 45–100%.

5 days ago·33 views

Materials & Energy

Tiny pores help organic frameworks turn otherwise lost energy into light

Nanopore confinement enabled delayed fluorescence in crystalline organic frameworks, allowing them to use energy that would otherwise be lost through non-emissive triplet states. The resulting materials formed high-efficiency light-emitting films and were used in solution-processed organic light-emitting diodes, although the abstract does not report detailed performance figures.

5 days ago·18 views

Abstract editorial illustration accompanying the article: Rotating discs helped a solar still make desalinated water for under 11 cents a liter

Materials & Energy

Rotating discs helped a solar still make desalinated water for under 11 cents a liter

A solar still fitted with rotating aluminum discs produced 24% more desalinated water than a drum-based still and 310% more than a standard passive still in experiments. Estimated water costs ranged from 8.2 to 10.3 cents per liter, with projected payback periods under seven months.

1 week ago·58 views