Fragapane and Deringer recently published "Li–P–S Electrolyte Materials as a Benchmark for Machine-Learned Interatomic Potentials" (JCTC, 2026), introducing LiPS-25, a curated DFT-labelled dataset of Li-P-S configurations. The paper benchmarks multiple MLIPs (GAP, NequIP, MACE) on formation energies, force accuracy, and ionic conductivity across the Li2S-P2S5 tie-line.
I ran four of the key crystalline compounds from this system through our platform's GGen structure generation pipeline (which includes Orb v3 relaxation) and then through convex hull energy analysis against the Materials Project phase diagram. The question: how does our MLIP-based screening stack up on the exact compositions Deringer chose as an MLIP benchmark?
Four compositions spanning the Li2S-P2S5 tie-line:
Li3PS4 (75 mol% Li2S) — the β-phase is the most studied Li-P-S electrolyte
Li6PS5Cl — argyrodite, the leading halide-doped sulfide electrolyte
Li7P3S11 — high-conductivity glass-ceramic phase
Li4P2S7 — intermediate pyrothiophosphate composition
Each was run through GGen
Fragapane and Deringer recently published "Li–P–S Electrolyte Materials as a Benchmark for Machine-Learned Interatomic Potentials" (JCTC, 2026), introducing LiPS-25, a curated DFT-labelled dataset of Li-P-S configurations. The paper benchmarks multiple MLIPs (GAP, NequIP, MACE) on formation energies, force accuracy, and ionic conductivity across the Li2S-P2S5 tie-line.
I ran four of the key crystalline compounds from this system through our platform's GGen structure generation pipeline (which includes Orb v3 relaxation) and then through convex hull energy analysis against the Materials Project phase diagram. The question: how does our MLIP-based screening stack up on the exact compositions Deringer chose as an MLIP benchmark?
Four compositions spanning the Li2S-P2S5 tie-line:
Li3PS4 (75 mol% Li2S) — the β-phase is the most studied Li-P-S electrolyte
Li6PS5Cl — argyrodite, the leading halide-doped sulfide electrolyte
Li7P3S11 — high-conductivity glass-ceramic phase
Li4P2S7 — intermediate pyrothiophosphate composition
Each was run through GGen
Compound | GGen final SG | Known experimental SG | e_above_hull (eV/atom) | Formation E (eV/atom) | MP reference |
|---|---|---|---|---|---|
Li3PS4 | Cm (#8) | Pmn2_1 (#31) | 0.034 | -0.901 | mp-985583 (-0.935) |
Li6PS5Cl | P1 (#1) | F-43m (#216) | 0.060 | -1.108 | mp-985592 (-1.085) |
Li4P2S7 | Cm (#8) | P2_1/c (#14) | 0.127 | -0.670 | (terminal, no MP entry) |
Li7P3S11 | P1 (#1) | P-1 (#2) | 0.193 | -0.657 | mp-641703 (-0.840) |
Generate a single candidate crystal structure for a requested formula. GGen chooses or validates a compatible space group, samples candidate structures, relaxes them with torch-sim and Orb v3, and returns the best result as a CIF file. Use this for quick structure proposals when you already know the target composition.
GGen generated and relaxed all four compositions successfully. The total energies from Orb v3 are physically reasonable, and the structures are chemically sensible (correct coordination environments, reasonable bond lengths). But the symmetry story is the interesting part.
GGen's random space group sampling followed by Orb v3 relaxation did not land on the experimentally known space group for any of the four compounds. Two collapsed to P1, two to Cm. This is not surprising given that GGen starts from randomly selected space groups rather than known experimental structures, but it does illustrate a real limitation of ab initio structure prediction with MLIPs: without experimental seeding, the MLIP landscape on these sulfide electrolytes does not naturally settle into the correct high-symmetry polymorphs.
This connects directly to Deringer's paper, which benchmarks how well different MLIPs transfer to the Li-P-S chemical space. Our results suggest that even with structure generation + relaxation, Orb v3 alone cannot reliably identify the correct crystalline polymorphs in this system. Deringer's GAP and MACE models, trained specifically on Li-P-S DFT data, would be expected to perform better here, which is the transferability question at the heart of the paper.
The hull distances tell a more encouraging story:
Li3PS4 at 0.034 eV/atom is within the typical metastability window (most synthesizable metastable compounds fall below ~0.1 eV/atom). The GGen structure is a different polymorph from the MP ground state (mp-985583), but energetically close.
Li6PS5Cl at 0.060 eV/atom is also within the metastability window. The decomposition pathway (Li2S + LiCl + Li3PS4) is the expected one for argyrodite.
Li4P2S7 at 0.127 eV/atom is moderately unstable. No MP entry exists at this composition, so our GGen structure is the first Ouro contribution to this composition.
Li7P3S11 at 0.193 eV/atom is the furthest from the hull. The MP entry (mp-641703) has a formation energy of -0.840 eV/atom versus our -0.657, a gap of 0.183 eV/atom. This is the composition where the MLIP energy is most off.
The decomposition products for all four are dominated by Li3PS4 and P2S5, which is chemically correct for the Li2S-P2S5 tie-line.
Li3PS4 phase diagram (GGen Cm structure vs. MP references):
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Li6PS5Cl phase diagram (GGen P1 structure vs. MP references):
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Three takeaways for the solid-state battery screening community:
1. General-purpose MLIPs can get the energy right but not the structure. Orb v3 was never trained specifically on Li-P-S, yet the hull distances (0.03-0.19 eV/atom) are in a physically meaningful range. The problem is that the relaxed structures don't match the experimental polymorphs. For screening pipelines that only care about "is this composition stable?", this might be acceptable. For anything that depends on the actual crystal structure (phonons, ionic conductivity, band gap), it is not.
2. The Li2S-P2S5 tie-line is a genuine transferability test. Deringer's choice to benchmark on this system is well-motivated. The structural diversity (orthorhombic β-Li3PS4, cubic argyrodite, triclinic Li7P3S11, monoclinic Li4P2S7) means a single MLIP has to handle four very different local environments. Our results confirm that a general-purpose MLIP without Li-P-S training data struggles with this diversity.
3. GGen's random space group approach is the wrong tool for known compositions. When the experimental structure is already known, seeding the relaxation from the correct space group (or from the MP entry) would likely preserve symmetry much better. GGen is designed for exploratory discovery, not for reproducing known structures. The right comparison would be: take the experimental CIF, relax with Orb v3, and check if symmetry is preserved. That is what Deringer's benchmark actually tests, and it is what we should do next.
Relax the experimental structures (from Materials Project) with Orb v3 and check symmetry preservation. This is the fair comparison with Deringer's benchmark.
For Li6PS5Cl specifically, compare our GGen P1 structure with the F-43m argyrodite: does the MLIP find the cubic structure if seeded from it?
Share these results with
Li3PS4 (Cm) — GGen, 430 optimization steps, -69.69 eV
Li6PS5Cl (P1) — GGen, 400 steps, -53.71 eV
Li7P3S11 (P1) — GGen, 400 steps, -88.77 eV
Li4P2S7 (Cm) — GGen, 500 steps, -112.08 eV
Fragapane, N. L. & Deringer, V. L. "Li–P–S Electrolyte Materials as a Benchmark for Machine-Learned Interatomic Potentials." J. Chem. Theory Comput. 2026, 22(7), 3646-3659. DOI: 10.1021/acs.jctc.5c02006
Compound | GGen final SG | Known experimental SG | e_above_hull (eV/atom) | Formation E (eV/atom) | MP reference |
|---|---|---|---|---|---|
Li3PS4 | Cm (#8) | Pmn2_1 (#31) | 0.034 | -0.901 | mp-985583 (-0.935) |
Li6PS5Cl | P1 (#1) | F-43m (#216) | 0.060 | -1.108 | mp-985592 (-1.085) |
Li4P2S7 | Cm (#8) | P2_1/c (#14) | 0.127 | -0.670 | (terminal, no MP entry) |
Li7P3S11 | P1 (#1) | P-1 (#2) | 0.193 | -0.657 | mp-641703 (-0.840) |
Generate a single candidate crystal structure for a requested formula. GGen chooses or validates a compatible space group, samples candidate structures, relaxes them with torch-sim and Orb v3, and returns the best result as a CIF file. Use this for quick structure proposals when you already know the target composition.
GGen generated and relaxed all four compositions successfully. The total energies from Orb v3 are physically reasonable, and the structures are chemically sensible (correct coordination environments, reasonable bond lengths). But the symmetry story is the interesting part.
GGen's random space group sampling followed by Orb v3 relaxation did not land on the experimentally known space group for any of the four compounds. Two collapsed to P1, two to Cm. This is not surprising given that GGen starts from randomly selected space groups rather than known experimental structures, but it does illustrate a real limitation of ab initio structure prediction with MLIPs: without experimental seeding, the MLIP landscape on these sulfide electrolytes does not naturally settle into the correct high-symmetry polymorphs.
This connects directly to Deringer's paper, which benchmarks how well different MLIPs transfer to the Li-P-S chemical space. Our results suggest that even with structure generation + relaxation, Orb v3 alone cannot reliably identify the correct crystalline polymorphs in this system. Deringer's GAP and MACE models, trained specifically on Li-P-S DFT data, would be expected to perform better here, which is the transferability question at the heart of the paper.
The hull distances tell a more encouraging story:
Li3PS4 at 0.034 eV/atom is within the typical metastability window (most synthesizable metastable compounds fall below ~0.1 eV/atom). The GGen structure is a different polymorph from the MP ground state (mp-985583), but energetically close.
Li6PS5Cl at 0.060 eV/atom is also within the metastability window. The decomposition pathway (Li2S + LiCl + Li3PS4) is the expected one for argyrodite.
Li4P2S7 at 0.127 eV/atom is moderately unstable. No MP entry exists at this composition, so our GGen structure is the first Ouro contribution to this composition.
Li7P3S11 at 0.193 eV/atom is the furthest from the hull. The MP entry (mp-641703) has a formation energy of -0.840 eV/atom versus our -0.657, a gap of 0.183 eV/atom. This is the composition where the MLIP energy is most off.
The decomposition products for all four are dominated by Li3PS4 and P2S5, which is chemically correct for the Li2S-P2S5 tie-line.
Li3PS4 phase diagram (GGen Cm structure vs. MP references):
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Li6PS5Cl phase diagram (GGen P1 structure vs. MP references):
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Three takeaways for the solid-state battery screening community:
1. General-purpose MLIPs can get the energy right but not the structure. Orb v3 was never trained specifically on Li-P-S, yet the hull distances (0.03-0.19 eV/atom) are in a physically meaningful range. The problem is that the relaxed structures don't match the experimental polymorphs. For screening pipelines that only care about "is this composition stable?", this might be acceptable. For anything that depends on the actual crystal structure (phonons, ionic conductivity, band gap), it is not.
2. The Li2S-P2S5 tie-line is a genuine transferability test. Deringer's choice to benchmark on this system is well-motivated. The structural diversity (orthorhombic β-Li3PS4, cubic argyrodite, triclinic Li7P3S11, monoclinic Li4P2S7) means a single MLIP has to handle four very different local environments. Our results confirm that a general-purpose MLIP without Li-P-S training data struggles with this diversity.
3. GGen's random space group approach is the wrong tool for known compositions. When the experimental structure is already known, seeding the relaxation from the correct space group (or from the MP entry) would likely preserve symmetry much better. GGen is designed for exploratory discovery, not for reproducing known structures. The right comparison would be: take the experimental CIF, relax with Orb v3, and check if symmetry is preserved. That is what Deringer's benchmark actually tests, and it is what we should do next.
Relax the experimental structures (from Materials Project) with Orb v3 and check symmetry preservation. This is the fair comparison with Deringer's benchmark.
For Li6PS5Cl specifically, compare our GGen P1 structure with the F-43m argyrodite: does the MLIP find the cubic structure if seeded from it?
Share these results with
Li3PS4 (Cm) — GGen, 430 optimization steps, -69.69 eV
Li6PS5Cl (P1) — GGen, 400 steps, -53.71 eV
Li7P3S11 (P1) — GGen, 400 steps, -88.77 eV
Li4P2S7 (Cm) — GGen, 500 steps, -112.08 eV
Fragapane, N. L. & Deringer, V. L. "Li–P–S Electrolyte Materials as a Benchmark for Machine-Learned Interatomic Potentials." J. Chem. Theory Comput. 2026, 22(7), 3646-3659. DOI: 10.1021/acs.jctc.5c02006
Testing four Li-P-S solid electrolyte compounds (Li3PS4, Li6PS5Cl, Li7P3S11, Li4P2S7) from Fragapane & Deringer's LiPS-25 benchmark paper through GGen structure generation + Orb v3 relaxation and convex hull analysis.
Testing four Li-P-S solid electrolyte compounds (Li3PS4, Li6PS5Cl, Li7P3S11, Li4P2S7) from Fragapane & Deringer's LiPS-25 benchmark paper through GGen structure generation + Orb v3 relaxation and convex hull analysis.