Three papers, fifteen years apart, that defined the field. A guided reading of what each one actually contributed, what specific problem it solved in its predecessor, and where to find them.
Behler & Parrinello, 2007
Schütt et al. (SchNet), 2017
Batzner et al. (NequIP), 2022
Understand each contribution in context
Reading papers in publication order — rather than starting from the newest — is the fastest way to actually understand why each architectural choice exists. Every later paper is solving a specific limitation of its predecessor. This post walks through that chain for the three papers most directly relevant to the equivariant networks covered in Post 17.
Read Behler-Parrinello first even though it's the oldest — every later paper assumes you understand "sum of per-atom neural networks" as the baseline architecture. Skipping straight to NequIP or MACE means encountering unexplained jargon (descriptors, message passing, equivariance) with no grounding in what problem they solve.
1. Behler & Parrinello (2007) — "Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces"
Physical Review Letters 98, 146401 (2007)
The problem it solved
Before 2007, anyone who tried using a single neural network to predict the total energy of a system directly from raw atomic coordinates ran into two hard walls: the network's output changed if you simply listed the atoms in a different order (no permutation invariance), and a network trained on a 10-atom system could not be applied to a 50-atom system (no size transferability).
The contribution
Gi = a fixed-length vector of "symmetry functions" — hand-designed,
rotation- and permutation-invariant descriptors of atom i's local environment
Ei = a small feedforward NN, with shared weights for every atom of the same element
By summing independent per-atom contributions, total energy automatically scales correctly with system size, and because every atom of the same element shares one network, the model immediately works on systems of any size — the network was effectively trained once per element, then reused atom by atom.
The symmetry functions Gi were hand-designed — a researcher had to choose the functional forms and parameters in advance, the same feature-engineering bottleneck Module 2 covers for tabular ML. The network could only ever work with whatever information the symmetry functions happened to preserve.
2. Schütt et al. (2017) — "SchNet: A Continuous-Filter Convolutional Neural Network for Modeling Quantum Interactions"
The problem it solved
Behler-Parrinello's hand-crafted symmetry functions were a fixed, finite basis — if the chosen functions didn't capture some relevant aspect of the local environment, no amount of training data could fix it. SchNet's central question: can the network learn its own descriptors directly from atomic distances, end-to-end, instead of being handed a fixed fingerprint?
The contribution
SchNet replaces hand-crafted symmetry functions with continuous-filter convolutions — learnable functions of interatomic distance that act like a convolution filter sliding over the local environment, but defined continuously (since atoms aren't on a fixed grid like image pixels). The filter weights are themselves learned during training, so the network discovers which distance patterns matter most for the property being predicted.
Wfilter is itself a small neural network mapping distance → filter weights
Stacking several interaction blocks lets information flow across multiple bonds
The result still keeps every guarantee from Behler-Parrinello (permutation invariance, size extensivity, rotation invariance) but lets the network learn what to look for rather than being told in advance.
SchNet's filters are still functions of distance alone — the architecture remains fully invariant and never represents bond direction explicitly. This is the exact limitation Post 17 covers, and the one NequIP was built to address.
3. Batzner et al. (2022) — "E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials" (NequIP)
Nature Communications 13, 2453 (2022)
The problem it solved
SchNet-style invariant networks needed large amounts of training data to learn accurate forces indirectly, because directional information had to be recovered through differentiation of distance-only features rather than being represented explicitly. NequIP's question: what if the network's internal features could be vectors and tensors from the start, so directional information never has to be reconstructed?
The contribution
NequIP builds every layer from E(3)-equivariant tensor operations — features are organised by rotation order ℓ (scalars, vectors, and higher-rank tensors simultaneously), and combined using Clebsch-Gordan tensor products that are mathematically guaranteed to preserve correct equivariance through arbitrarily many layers.
Yℓ' = spherical harmonics of the bond direction, up to some maximum order
The tensor product combines features of order ℓ' and ℓ'' into output order ℓ
Clebsch-Gordan coefficients ensure the combination is exactly equivariant
The headline result in the paper: NequIP matched or exceeded the accuracy of invariant networks trained on far more data — in some benchmarks, using 10-100× fewer training structures for the same force accuracy, because the model no longer had to "rediscover" directional information from scratch for every new chemistry.
DFT calculations on transition-metal chalcogenides are expensive — especially with DFT+U or hybrid functionals needed for correlated systems. NequIP's data efficiency result means a useful potential can in principle be trained on hundreds rather than tens of thousands of structures — directly relevant when your own calculation budget is the bottleneck, as discussed in Post 21's roadmap.
4. Where MACE Fits
MACE (Batatia et al. 2022, arXiv:2206.07697) was not covered as a full separate reading here because Post 17 already detailed its specific contribution: constructing higher-order many-body equivariant features within a single layer, rather than building them up gradually across many NequIP-style layers — making MACE both more expressive per parameter and substantially faster to train, which is part of why it was chosen as the backbone for the MACE-MP-0 foundation model used throughout this module.
5. Summary Table
| Paper | Solved | Left unsolved |
|---|---|---|
| Behler-Parrinello (2007) | Permutation invariance, size extensivity — via sum of per-atom NNs | Descriptors are hand-crafted, fixed in advance |
| SchNet (2017) | Learnable, end-to-end descriptors via continuous-filter convolutions | Still distance-only — no explicit directional (equivariant) features |
| NequIP (2022) | Full E(3) equivariance — tensor features rotate correctly with the crystal | Many-body correlations built up slowly, layer by layer |
| MACE (2022) | Higher-order many-body features within a single equivariant layer | (Current state of the art for most materials applications) |
Quick Check
1. What specific problem in pre-2007 neural-network energy models did Behler & Parrinello's architecture solve?
2. What did SchNet change relative to Behler-Parrinello, and what did it leave unchanged?
3. What was NequIP's headline empirical result compared to invariant (SchNet-style) networks?
My name is Dr Abderrahmane Reggad. I'm a 54 year old and I am a researcher in materials' sciences. 
