Review Papers Claim by Claim

Concise verdict

PiNN introduces PiNet2, an equivariant upgrade to PiNet that adds P3 vector and P5 tensor features and specialized modules PiNet2 dipole and PiNet2 chi; the authors show consistent improvements in energy force dipole and polarizability predictions across QM9 rMD17 Materials Project LIQWAT NaCl and protic ionic liquid benchmarks and provide plug ins PiNNAcLe and PiNNwall for adaptive on the fly learning and electrode simulations PiNN

Detailed critical review and analysis

Paper PiNN PiNet2 Equivariant Neural Network Suite for Modeling Electrochemical Systems DOI 10.1021/acs.jctc.4c01570 (Journal of Chemical Theory and Computation 2025)

What the authors did

• Extended the original invariant PiNet architecture to an equivariant PiNet2 with P3 (vectorial) and P5 (tensorial) features and implemented PiNet2 P3 and PiNet2 P5 variants PiNet2
• Built specialized modules PiNet2 dipole (for dipoles and base charges) and PiNet2 chi (for atom condensed charge response kernels and polarizability) and integrated adaptive learn on the fly PiNNAcLe and electrode plugin PiNNwall for coupling to MetalWalls PiNNwall

Datasets and benchmarks

The authors benchmarked across representative public datasets and realistic case studies:

• QM9 purified ~134k molecules (dipole energy) rMD17 small molecules (forces) QM9
• Materials Project crystals MP crystals 2018.6.1 and MPF.2021.2.8 (M3GNet dataset) for solid materials energy and forces
• Liquid datasets LIQWAT H2O revPBE0 D3 NaCl(sol) SCAN and a protic ionic liquid [C1M][HOAc] case for transport and reactive MD

Key reported results (verbatim metrics from paper)

Authors claims vs evidence

• Claim PiNet2 P3 gives universal improvement with modest overhead. Evidence improvement is sizable in low symmetry liquid and ionic data sets (H2O NaCl protic ionic liquid) and smaller for high symmetry crystals where improvements were marginal; authors show timing/overhead in Supporting Information and recommend PiNet2 P3 as cost effective versus P5 Computational
• Claim PiNet2 modules enable electrode simulations under external bias via PiNNwall and MetalWalls. Evidence authors present graphene oxide/electrolyte interface simulations PZEC and capacitance trends with base charge shifts but show Helmholtz capacitance results require careful conversion (factor of 2 mistake acknowledged) — transparency is good but independent replication is warranted PiNNwall

Strengths

• Comprehensive benchmarking across molecules solids and liquids with clear numeric metrics and supporting information enabling reproduction of training setups and hyperparameters Supporting
• Clear demonstration that equivariance helps liquids and disordered systems where angular information matters strongly (water structure densities conductivities) with mechanistic diagnostics (RDF density equilibrium properties IR spectra) Equivariance

Limitations and blind spots

• Dependence on DFT references and functional choice: authors used several functionals (revPBE0 D3 SCAN B3LYP) and show some properties (e g ionic conductivity fractions proton transfer degrees) depend on the reference functional; results therefore inherit functional biases and need experimental cross checks Functional
• Computational overhead for PiNet2 P5 is significant and authors advise PiNet2 P3 as cost effective compromise but exact runtime tradeoffs depend on hardware and production pipelines (Supporting Info contains S1 figure) PiNet2
• Generality to arbitrary electrode chemistries and extreme conditions is not yet proven; tests shown focus on graphene oxide and protic ionic liquids — broader validation required
• Some reported observables (e g Helmholtz capacitance numerical factor) required correction and authors transparently flagged the conversion factor issue — users must carefully follow SI to reproduce numbers Helmholtz

Reproducibility and code availability

Code is public and documented at the project GitHub repository and supporting information contains hyperparameters training splits and timing benchmarks which supports reproducibility though reproducing large MLP MD studies requires substantial compute Code

Repository top level README indicates installation instructions requirements and links to docs and examples which is positive for adoption and verification PiNN

Practical recommendations for users

1. Use PiNet2 P3 for liquids ionic solutions and low symmetry systems where angular information is important; expect measurable gains in forces and thermodynamic properties (density RDF) PiNet2
2. Reserve PiNet2 P5 only when highest-fidelity angular/tensor features justify the extra cost and hardware constraints permit it
3. Cross validate MLP predictions against at least two DFT functionals if properties are sensitive to exchange correlation functional choice (e g transport proton transfer) Functional

Where this could be refuted

Show that PiNet2 P3 does not generalize outside the covered chemical space by testing substantially different electrolytes electrode materials or reactive chemistries and demonstrate no improvement in force/energy/RDF/density or worse MD stability; or repeat PiNNwall electrode MD with independent codes and show systematic disagreement.

Conclusion

PiNN PiNet2 is a high quality careful contribution that demonstrates the practical value of equivariant features for ML potentials focused on electrochemical modeling The paper is well documented reproducible and provides useful plugins for on the fly learning and electrode coupling The principal caveats are dependence on DFT training references and computational overhead for higher order features Overall this is a solid technical advance with high practical utility for the field Paper