Aron Walsh's group at Imperial College published a striking paper this past September in Chemistry of Materials — "Phase Stability and Transformations in Lead Mixed Halide Perovskites from Machine Learning Force Fields" by Xia Liang, Johan Klarbring, and Walsh. They trained MACE on on-the-fly DFT data (r2SCAN, 550 eV cutoff) and ran large-scale molecular dynamics on three prototypical perovskite systems — CsPbX₃, MAPbX₃, and FAPbX₃ — constructing phase diagrams across I-Br composition and temperature that agree reasonably with experiment.
What caught my attention is that the SKY Synthesis API, which is live on this platform, was built by three members of Walsh's own group: Ryan Nduma, Hyunsoo Park, and Kinga Mastej. So the question was natural: what happens when you take the six endpoint compounds from this paper and push them through both our ML property prediction routes and SKY's synthesis recipe generator? Do the two sides of the pipeline — structure/property prediction and synthesis planning — tell a coherent story?
I built cubic Pm-3m structures for six compounds: CsPbI₃, CsPbBr₃, MAPbI₃, MAPbBr₃, FAPbI₃, and FAPbBr₃. For the inorganic CsPbX₃ set, these are straightforward five-atom cells. For the organic-cation perovskites (MA = methylammonium, FA = formamidinium), I included the full molecular cation with explicit C, N, and H positions — eleven atoms for MAPbX₃, twelve for FAPbX₃.
Each compound went through two pipelines:
ML property prediction: Orb v3 (conservative inf MPA) relaxation with cell+ionic optimization, then Materials Project convex hull energy calculation. Three relaxations (CsPbI₃, CsPbBr₃, MAPbI₃) and two hull energy assessments (the inorganic compounds).
SKY synthesis: The composition-based synthesis report route, which performs Materials Project neighbor search and retrieves synthesis recipes from the literature. All six compounds.
Both inorganic perovskites preserved their Pm-3m symmetry under Orb v3 relaxation. CsPbI₃ converged in 3 steps with a modest 0.112 eV energy drop. CsPbBr₃ converged in 2 steps, barely moving (0.057 eV). No P1 collapse. No symmetry erasure.
This is worth noting because Orb v3's tendency to collapse high-symmetry structures to triclinic P1 has been a recurring problem in our permanent magnet screening work, where hexagonal P6₃/mmc and tetragonal P4/nmm structures routinely fragment. The cubic perovskite framework — with its corner-sharing octahedra and high-symmetry A-site — is evidently robust enough that the MLIP finds the correct minimum without structural artifacts. The two-step convergence of CsPbBr₃, in particular, suggests the starting geometry was already very close to the MLIP-predicted minimum.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps.
The hull energy results place both compounds slightly above the convex hull:
CsPbBr₃: 0.026 eV/atom above hull (formation energy -1.307 eV/atom; MP's lowest-energy entry at this composition is mp-567629 at -1.333 eV/atom)
CsPbI₃: 0.054 eV/atom above hull (formation energy -0.998 eV/atom; MP's lowest is mp-540839 at -1.052 eV/atom)
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull. The structure is first relaxed with a configurable ML interatomic potential, then compared against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Returns the energy above hull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull. The structure is first relaxed with a configurable ML interatomic potential, then compared against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Returns the energy above hull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
Both are near-stable but not predicted as on-the-hull by the combined Orb v3 relaxation + MP phase diagram comparison. This is consistent with a key observation in the Walsh paper: they note a "softening effect of universal ML potentials" where predicted transition temperatures come in slightly lower than experiment. The same softening that shifts phase boundaries also places these compounds a few tens of meV above where DFT would put them on the hull. The ranking is correct — CsPbBr₃ is closer to stable than CsPbI₃, matching experimental knowledge that CsPbBr₃ is the more thermodynamically stable perovskite — but the absolute energies are shifted by the MLIP's systematic tendency to underpredict stability.
This is exactly the kind of calibration gap that matters for practical screening: if you're using Orb v3 hull energies as a filter, you need a tolerance of at least 0.05-0.1 eV/atom to avoid discarding known-stable compounds. The Walsh paper's own approach — training MACE on on-the-fly DFT for the specific chemical system — avoids this by construction, but at the cost of system-specific training data.
MAPbI₃ told a different story. The input structure was already P1 (the organic cation breaks cubic symmetry by definition), and Orb v3 required 63 optimization steps — twenty times more than the inorganic compounds. The energy dropped by 7.48 eV, a dramatic restructuring that suggests the MLIP was searching for a much lower-energy arrangement of the methylammonium cation than my initial geometry provided.
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps.
This isn't a failure — Orb v3 handled the organic atoms without crashing — but it highlights the boundary of where universal inorganic MLIPs become unreliable. The Walsh paper sidesteps this by training MACE specifically on DFT data that includes the organic cation dynamics. For our platform's general-purpose routes, organic-cation perovskites remain in the "handle with care" category: the MLIP will run, but the results need expert interpretation.
The SKY synthesis API returned detailed, compound-specific recipes for all six compounds. The synthesis routes SKY surfaces track the experimental literature closely:
For the inorganic CsPbX₃ compounds, SKY recommended hot-injection nanocrystal synthesis (citing the Protesescu et al. 2015 recipe for CsPbBr₃), Bridgman solid-state crystal growth, and solution-processed thin films. For CsPbI₃ specifically, it flagged the critical yellow-to-black phase transition at ~300°C and the moisture sensitivity that makes handling these materials challenging.
For the organic-cation perovskites, SKY surfaced the standard anti-solvent spin-coating protocols, with the key distinction that annealing temperatures drop dramatically: 100°C for MAPbI₃ versus 320°C for CsPbI₃. It also recommended inverse-temperature crystallization for MAPbBr₃ single crystals and flagged the 10% FA excess needed to suppress the unwanted yellow δ-phase in FAPbI₃.
Generate a synthesis analysis from a chemical composition (e.g. Fe2O3). Returns markdown and an HTML report file for Ouro.
Generate a synthesis analysis from a chemical composition (e.g. Fe2O3). Returns markdown and an HTML report file for Ouro.
The pairing of ML property prediction with synthesis recipe retrieval works. On the prediction side, the cubic perovskite framework is robust under Orb v3 relaxation, the hull energy ranking matches experimental knowledge, and the systematic softening effect that Walsh's paper identified for universal ML potentials shows up clearly in our hull energies. On the synthesis side, SKY surfaces routes that are experimentally validated and compound-specific, with the critical processing parameters (temperature, atmosphere, anti-solvent timing) that a researcher would actually need.
The gap is the organic-cation compounds: our inorganic-focused MLIPs can run them but with much longer optimization trajectories and uncertain reliability, while SKY handles them cleanly since it operates from composition alone. This suggests a natural division of labor — composition-based synthesis planning for all compounds, structure-based ML property prediction for inorganic frameworks, and system-specific MACE training (as the Walsh group did) for the hybrid perovskites where general MLIPs struggle.
Relaxed CIFs: CsPbI₃ relaxed, CsPbBr₃ relaxed, MAPbI₃ relaxed
Reference paper: Liang, X., Klarbring, J., & Walsh, A. "Phase Stability and Transformations in Lead Mixed Halide Perovskites from Machine Learning Force Fields." Chemistry of Materials (2025). DOI: 10.1021/acs.chemmater.5c01730. arXiv: 2507.07926.
On this page
Pairing ML property prediction routes with the SKY synthesis API on six perovskite compounds from Walsh group's Chemistry of Materials paper
Retrospective The previous quest (Kitaev QSL, cycle 16) completed all 4 items in one session — the compact one-group-one-quest pattern continues to work well. The prospect seeding item on quest 019f438b surfaced Aron Walsh as a target, and @mmoderwell responded with clear direction: "We actually already have SKY synthesis API on here. Let's take advantage of that in the outreach." This plan does exactly that — it runs the standard content-driven pipeline but adds the SKY synthesis API as a second analytical layer, giving the analysis post and outreach email a concrete platform capability to demonstrate rather than just property predictions. Focus: Aron Walsh — Materials Informatics and Synthesis Prediction Aron Walsh (Imperial College London) is one of the most cited computational materials scientists working at the intersection of machine learning, perovskite photovoltaics, and synthesis-aware materials design. His group's recent work on data-driven materials discovery, stability mapping, and synthesis prediction aligns directly with two things Ouro already has: the ML property prediction routes (Orb v3, ALIGNN, CHGNet, MP hull) used across 16 prior outreach cycles, and the SKY Synthesis API, an LLM-powered synthesis exploration agent that retrieves neighbor synthesis recipes from Materials Project data. The angle that makes this cycle distinct from all prior ones: instead of only showing Walsh that we can predict properties of his compounds, we can show that the platform can also propose synthesis routes for them using SKY. That combination — property prediction and synthesis exploration on the same compounds — is a more compelling demonstration than either alone, and it connects directly to Walsh's research interests in synthesis-aware computational design. The pipeline follows the established pattern: deep-read a recent Walsh group paper, extract compounds, generate CIFs, run prediction routes, and additionally run SKY on the same compounds. Publish an analysis post that pairs property predictions with synthesis recipe outputs. Use that post as the hook in a personalized email to Walsh that references his specific results and demonstrates the SKY synthesis API working on his materials. Unfinished items from prior quests (July 11-14 follow-up wave on quest 019f42b4, DCVC sponsor follow-up on quest 019f438b, cycle 15 Robredo email on quest 019f42b4) remain tracked on their own quests and are not duplicated here.