A chemical separation begins with a preference: a molecule must interact differently with the metals mixed together. Researchers exploring rare-earth extraction therefore have a molecular design problem as well as a processing problem. Which chemical grip is worth making and testing?[3]
Berkeley Lab researchers investigated that question with Allegro, a neural-network architecture that respects molecular geometry. Their Digital Discovery paper, published online in November 2025, trained models on calculated energies for 5,356 rare-earth complexes.[1]
The strongest result concerns what the model should learn. Predicting binding energy directly gave a mean absolute error of 6.1 kilocalories per mole on the held-out computed test set. Predicting total energy first, then deriving binding energy using calculated fragment energies, produced an error of 11.3. These measure agreement with quantum-chemical calculations, not extraction yields.[1]
A difference can be harder than a total
Our arithmetic illustration shows why the distinction matters. Suppose a complex has a total energy of −10,000 arbitrary units and its separate parts total −9,800. Their difference is −200. An error of 10 units in the complex’s energy is just 0.1% of its magnitude, yet changes the difference by 5% if the parts are exact. An impressive total-energy prediction can therefore be much less impressive for the quantity a chemist actually needs.
That example explains the arithmetic; it does not establish why one trained network outperformed the other. The practical design choice is to check accuracy on the target quantity itself.
Allegro’s underlying 2023 methods paper explains another useful constraint. The energy of a molecular structure should stay the same when the entire structure is rotated or translated. Quantities such as force vectors must transform with the geometry. The architecture builds these symmetry relationships into local atomic representations, so its internal calculations preserve the appropriate physical relationships.[2]
Think of rotating a complete molecular model on a screen: the camera angle has changed, but no chemical bond has. Changing a group attached to the molecule is a different operation. A useful predictor must ignore the former while remaining sensitive to the latter.
The liquid still has the last word
The direct-binding model also retained useful accuracy on 325 complexes built with a previously unseen ligand isomer, giving a mean absolute error of 9.4 kilocalories per mole. The study provides a screening approach, rather than a demonstrated industrial separation process.[1]
A separate benchtop study published in 2020 shows what an experimental separation has to establish. Researchers examined a ligand binding lanthanum, neodymium and dysprosium, varied acidity, filtered out precipitated complexes, and measured the metal remaining in solution. Changing conditions changed which complex precipitated. That experiment was not a validation of the newer AI model.[3]
Binding predictions can help prioritise candidate molecules. The wet experiment asks whether a candidate actually divides a mixture under workable conditions. Connecting those two tasks is how a better molecular estimate becomes useful chemistry.
Sources & further reading
- Toward accelerating rare-earth metal extraction using equivariant neural networks Digital Discovery · 26 Nov 2025 · Accessed 20 Oct 2026
- Learning local equivariant representations for large-scale atomistic dynamics Nature Communications · 3 Feb 2023 · Accessed 20 Oct 2026
- High-throughput screening for discovery of benchtop separations systems for selected rare earth elements Communications Chemistry · 20 Jan 2020 · Accessed 20 Oct 2026