ML-guided screening of chalcogenide perovskites as solar energy materials

Abstract: Chalcogenide perovskites have emerged as promising absorber materials for next-generation photovoltaic devices, yet their experimental realization remains limited by competing phases, structural polymorphism, and synthetic challenges. Here, we present a fully data-driven and experimentally grounded screening and ranking framework to assess the stability and experimental feasibility of chalcogenide perovskites, integrating interpretable analytical descriptors, machine-learning models, and sustainability metrics. Using a curated experimental dataset of halide and chalcogenide compounds, we derive a new tolerance factor via the SISSO (sure independence screening and sparsifying operator) algorithm that more accurately distinguishes perovskite-forming compositions than established tolerance-factor-based screening criteria. This descriptor is combined with generative crystal structure prediction, composition-based bandgap estimation, and machine-learning-based feasibility assessment to systematically explore a wide chemical space of hypothetical chalcogenide perovskites. The resulting candidates are further evaluated using sustainability indicators, enabling multi-objective ranking tailored to both single-junction and tandem photovoltaic architectures. Beyond identifying several promising and previously unexplored chalcogenide perovskites, this work demonstrates a transferable screening strategy for chemically constrained materials spaces that balances optoelectronic performance, experimental viability, and long-term sustainability.

• The paper introduces a SISSO-derived tolerance factor that predicts chalcogenide perovskite stability with 91.2% accuracy, 88.4% F1, and 95.0% precision, outperforming established factors.
• The workflow combines CrystaLLM structure generation and connectivity filtering with CrabNet bandgap prediction, narrowing 1,392 compositions to 54 corner-sharing candidates with a mean predicted bandgap of 1.86 eV.
• The paper ranks candidates using bandgap suitability, supply risk, and crystal-likeness, highlighting BaZrS3 for tandem photovoltaics and unexplored targets such as CuHfS3 and CuZrSe3 while emphasizing uncertainty for selenides.

Overview

This paper presents a machine-learning-guided screening pipeline for identifying synthesizable chalcogenide perovskite absorbers for photovoltaics (2602.21812). The workflow integrates four partially independent criteria: a newly derived SISSO-based tolerance factor $\tau^*$ for structural stability, generative crystal structure prediction via CrystaLLM, composition-based bandgap estimation with CrabNet, and sustainability metrics based on supply risk. The pipeline reduces a chemical space of 1392 charge-balanced ABX$_3$ chalcogenide compositions to 181 geometrically stable candidates, then to 54 compounds whose generated structures adopt true corner-sharing $BX_6$ networks, and finally to roughly 30 high-synthesizability candidates ranked by multi-objective criteria.

ML-derived tolerance factor

The central methodological contribution is the SISSO-derived descriptor $\tau^*$, trained on 283 experimentally reported halide and chalcogenide ABX$_3$ compounds (with Turnley et al. chalcogenide-tailored ionic radii). Perovskite stability is defined by $\tau^* < 0.846$, learned via a depth-one decision tree. On the held-out test set, $\tau^*$ achieves 91.2% accuracy and an F1-score of 88.4%, substantially outperforming the Goldschmidt factor (59.6% accuracy), and both the Jess et al. and Bartel factors (70.2% accuracy each).

The most important distinction is in precision: $\tau^*$ produces only one false positive versus many for $t_{\text{Jess}}$ and $t_{\text{Bartel}}$, yielding 95.0% precision and 97.1% specificity at identical recall (82.6%). Since false positives directly translate into wasted synthesis effort on compositions that favor competing non-perovskite phases—a well-documented failure mode in this materials family—this precision gain is the practically relevant improvement. Structurally, $_3$0 is asymmetric between sites: the A-site radius enters through a dominant cubic term $_3$1 imposing a sharp upper bound, while the B-site radius contributes mainly through a logarithmic size-mismatch term. Within the stability region ($_3$2 branch), $_3$3 governs feasibility while $_3$4 acts as a fine-tuning parameter. The authors note explicitly that the choice of ionic radii dataset is a modeling assumption whose refinement could alter the quantitative form of the descriptor.

Applying $_3$5 to the enumerated space yields 181 predicted stable compositions—only ~13% of the space—consistent with the experimentally observed scarcity of chalcogenide perovskites.

Structural validation via generative structure prediction

Each of the 181 stable compositions was passed through CrystaLLM, an autoregressive LLM trained on over three million DFT-derived CIFs, generating four candidate cells per formula. Structures were filtered by analyzing octahedral connectivity, since edge-sharing motifs within Pnma mimic corner-sharing perovskites. This step reduced the pool from 181 to 54 compositions (~30% of the $_3$6-stable subset). The authors frame this as a deliberate trade-off favoring precision over recall, motivated by the experimental prevalence of non-perovskite byproduct phases.

The refined set reveals a narrower stability window than $_3$7 alone: the A-site ratio concentrates at $_3$8, while the B-site ratio is confined to $_3$9, indicating that only a narrow octahedral-size range supports stable corner-sharing connectivity. Space-group analysis shows Pnma dominance (116 of 181), followed by Cmcm (46). Parity comparisons against ICSD data for known chalcogenide perovskites show close agreement in lattice parameters and volume. Importantly, DFT relaxation was intentionally omitted—CrystaLLM outputs are screening-level indicators, not definitive structural assignments, and subtle distortions or alternative low-energy phases may be missed. Notably, the validated set recovers known compounds (AScS$BX_6$0, BaHfS$BX_6$1/BaZrS$BX_6$2/BaUS$BX_6$3, EuHfS$BX_6$4/EuZrS$BX_6$5, SrSnS$BX_6$6) while predicting new candidates including several AZrS$BX_6$7 (A = Dy, Gd, La, Sm) rare-earth sulfides and EuB S$BX_6$8 variants.

A limitation worth flagging: the training set contains only one experimentally reported selenide perovskite among 27 chalcogenide examples (~5% of the test set), so all ABSe$BX_6$9 predictions carry large intrinsic uncertainty, as the authors acknowledge.

Bandgap estimation

Bandgaps were predicted with CrabNet trained on a curated dataset of 3628 unique compositions assembled from hybrid halide perovskites, chalcogenide perovskites, and chalcogenide semiconductors. The held-out test MAE is 248 meV ($\tau^*$0), within the 0.3–0.5 eV range typically reported for composition-only Matbench-style benchmarks, though the authors caution this comparison is contextual rather than strict given dataset dependence. Encoder ablations (Magpie, mat2vec, one-hot, Pettifor, random embeddings) show near-invariant performance, indicating the attention mechanism learns chemically meaningful representations regardless of initialization priors.

Two caveats temper these numbers. First, LOOCV restricted to the nine chalcogenide perovskite compositions yields a 281 meV MAE—only marginally better than trivial median/mean baselines (308/314 meV)—so extrapolation to the target class is only weakly supported. Second, LaScS$\tau^*$1's predicted gap of $\tau^*$2 eV underestimates a recent experimental value of 2.9 eV despite inclusion in training, although the paper notes strong absorption near 2 eV in the reported spectra and absence of measurement uncertainty in that study. PCA of the predictions attributes over 60% of bandgap variance to B-site cation and anion descriptors, consistent with the established picture of X-p / B-d derived band edges. The CrystaLLM-validated subset exhibits a mean predicted bandgap of 1.86 eV, naturally aligning these materials with tandem top-cell requirements rather than single-junction optimums.

Sustainability and multi-objective ranking

Supply risk was computed per compound from elemental Herfindahl-Hirschman indices (USGS 2025 production shares) weighted by World Bank-derived ESG scores, following Nominé et al. Uranium-containing candidates were excluded on toxicity grounds. Pareto analysis against the Shockley–Queisser-optimal bandgap deviations identifies BaZrS$\tau^*$3 as optimal for tandem configurations ($\tau^*$4 eV)—consistent with its status as the field's benchmark—and surfaces previously unexplored candidates CuHfS$\tau^*$5, CuZrSe$\tau^*$6, and EuYbSe$\tau^*$7. A broader near-Pareto set includes rare-earth-rich compositions such as EuYbS$\tau^*$8, CeScS$\tau^*$9, and EuScS$_3$0, though their elevated supply risk limits deployment-scale viability.

Experimental plausibility was assessed with the pre-trained GCNN of Gu et al., which outputs a crystal-likeness score (CLS) from positive-unlabeled learning; no fine-tuning was performed. All experimentally reported chalcogenide perovskites show CLS > 0.88, providing an empirical threshold above the generic 0.5 cutoff. High-CLS trends correlate with sulfides, Ba/La A-sites, and Sc/Tm B-sites. A Spearman correlation analysis across the 54 candidates confirms weak pairwise correlations among all four metrics ($_3$1), supporting the claim that the screening stages contribute complementary rather than redundant information.

Limitations and open questions

The paper concedes several substantive limitations at the points where they bear on results: (i) selenide perovskites are severely underrepresented in all training datasets, making ABSe$_3$2 predictions statistically fragile; (ii) CrystaLLM structures are unrelaxed and serve only as screening indicators; (iii) the LOOCV bandgap margin over null baselines for chalcogenide perovskites specifically is small (281 vs. 308 meV); (iv) redox stability is not modeled—pymatgen oxidation-state assignments may misrepresent charge-balance reality, e.g., Cu$_3$3 susceptibility to reduction in soft-anion environments per HSAB arguments, and Ce$_3$4 instability in sulfides; and (v) CLS is a statistical similarity metric blind to kinetic barriers and synthesis pathways. Whether any of the newly proposed compositions—EuScS$_3$5, LaTbS$_3$6, EuYbSe$_3$7, CuHfS$_3$8, CuZrSe$_3$9 among them—can actually be synthesized, and whether their predicted bandgaps survive experimental measurement, remains the open question the ranking is designed to answer.

Conclusion

This work demonstrates a transferable, interpretable multi-stage screening strategy for chemically constrained materials spaces, anchored by a SISSO tolerance factor with markedly improved precision over classical descriptors. Its strength lies less in any single model than in the demonstrated orthogonality of its four criteria, which together convert a 1392-compound hypothetical space into a shortlist of prioritized, sustainability-aware experimental targets.