Perovskite-R1: LLM for PSC Additive Design
- Perovskite-R1 is a domain-specific LLM that extracts actionable precursor-additive insights from 1,232 publications and a 33,269-compound library.
- It employs instruction-tuning with explicit chain-of-thought reasoning to guide additive selection for defect passivation, crystallization control, and energy-level alignment.
- Fine-tuned from a QwQ-32B base using LoRA, the model achieves a ~10% performance gain on challenging perovskite tasks, enhancing experimental design protocols.
Perovskite-R1 is a domain-specialized LLM for precursor-additive discovery and experimental design in perovskite solar cells. It was introduced as a system that mines 1,232 peer-reviewed publications together with a 33,269-compound candidate library, converts that material into an instruction-tuning corpus with explicit chain-of-thought reasoning, and fine-tunes a QwQ-32B base model to generate additive recommendations for defect passivation, crystallization control, energy-level alignment, and process compatibility (Wang et al., 22 Jul 2025). Its emergence reflects a broader PSC landscape in which rapid efficiency gains coexist with persistent bottlenecks in stability, environmental robustness, and scalable processing, as seen in bandgap-optimized all-perovskite tandems and in composition-dependent stability trade-offs in mixed-halide absorbers (Eperon et al., 2016, Yu et al., 2019).
1. Research setting and problem definition
Perovskite-R1 is positioned within a PSC literature in which best-in-class power conversion efficiencies have reached 26.95% since 2009, while commercialization remains constrained by long-term operational stability, environmental sustainability, and scalable manufacturing (Wang et al., 22 Jul 2025). In this setting, precursor additive engineering is treated as a practical intervention because it can improve stability by passivating intrinsic defects and controlling crystallization, enhance environmental sustainability by enabling benign additives and solvents and by mitigating degradation pathways, and support scalability through solution-processing compatibility and film-quality control (Wang et al., 22 Jul 2025).
The system addresses a specific epistemic bottleneck rather than a new absorber chemistry. The perovskite literature spans structurally and electronically diverse materials classes, including mixed Sn–Pb tandem absorbers, mixed-halide inorganic alloys, and lead-free double or perovskite-related compounds whose indirect gaps, broad photoluminescence, or highly insulating transport complicate direct transfer of design rules (Eperon et al., 2016, Lozhkina et al., 2017, Sun et al., 2015, Wei et al., 2016). Perovskite-R1 therefore targets the problem of systematically extracting actionable additive knowledge from a rapidly growing and chemically heterogeneous corpus rather than relying on manual screening, which the authors describe as slow and susceptible to cognitive bias and incomplete coverage (Wang et al., 22 Jul 2025).
Its stated scope is narrower than the entire perovskite field. The model focuses on defect passivation strategies in 3D lead-halide perovskites, additive selection from a large drug-like candidate library, and synthesis and processing guidelines such as solubility in DMF/DMSO, crystallization control, energy-level alignment, and compatibility with ETL/HTL stacks, all translated into practical protocols through structured prompts (Wang et al., 22 Jul 2025).
2. Corpus construction and candidate-space design
The knowledge base behind Perovskite-R1 combines literature curation with compound-space expansion. The literature component contains 1,232 peer-reviewed research articles curated from ACS, Elsevier, RSC, Wiley, and Springer Nature, with selection prioritizing detailed precursor-additive synthesis pathways, material characterization, performance optimization strategies, significant achievements in device efficiency or stability, and temporal breadth from early additive strategies to recent breakthroughs (Wang et al., 22 Jul 2025). Coverage is described as holistic across materials chemistry, processing, and device architecture, although explicit counts per domain are not reported (Wang et al., 22 Jul 2025).
The candidate-material component consists of a drug-like molecular library of 33,269 distinct compounds (Wang et al., 22 Jul 2025). Compounds are identified by common names and CAS registry numbers rather than by SMILES or InChI, and in the Methods section they are organized into seven major therapeutic categories, yielding 4,753 individual data records used in instruction-data generation (Wang et al., 22 Jul 2025). The paper emphasizes functional-group reasoning rather than a fixed additive taxonomy: the model analyzes motifs such as carboxyl, hydroxyl, pyridine-N, and phosphine oxide for coordination and passivation potential (Wang et al., 22 Jul 2025).
Preprocessing converts PDFs to text with pdfminer, then segments each paper into chunks of up to 2,500 tokens with 20% overlap to preserve semantic continuity (Wang et al., 22 Jul 2025). OpenAI o1 is then used to generate high-quality question–answer pairs with explicit chain-of-thought reasoning guided by a customized template (Wang et al., 22 Jul 2025). Literature chunks yield mechanistic and design Q&A, while compound entries yield stepwise functional-group analyses and suitability assessments (Wang et al., 22 Jul 2025). The resulting ontology centers on defect passivation rationales, crystallization control, energy-level alignment, solvent and process compatibility, and additive ranking (Wang et al., 22 Jul 2025).
Two qualifications are important. First, the paper does not report human validation of chain-of-thought steps, although benchmark questions were independently assessed by domain experts (Wang et al., 22 Jul 2025). Second, deduplication and quality-check specifics are not detailed, so the curation pipeline is richly described at the level of sources, chunking, and prompt generation, but less so at the level of post-generation filtering (Wang et al., 22 Jul 2025).
3. Base model, architecture, and fine-tuning regime
Perovskite-R1 is built by fine-tuning QwQ-32B, a Qwen2.5-32B-based causal LLM with approximately 32.5B parameters, including 31B non-embedding parameters and 64 Transformer layers (Wang et al., 22 Jul 2025). The base model uses Rotary Position Embedding, SwiGLU activations, RMSNorm, attention QKV bias, and Grouped-Query Attention with 40 query heads and 8 shared key–value heads, supporting a context length of up to 131,072 tokens (Wang et al., 22 Jul 2025). These design choices place the system in the class of long-context reasoning models rather than lightweight retrieval front ends.
Fine-tuning is parameter-efficient rather than full-model. The paper reports LoRA across all weight matrices with rank $16$, alpha $32$, and dropout $0.1$, trained in bfloat16 with FlashAttention2 for attention efficiency (Wang et al., 22 Jul 2025). The per-device batch size is $1$, gradient accumulation is $8$, the learning rate is with cosine decay, warmup is , and the total run lasts 10 epochs (Wang et al., 22 Jul 2025). The training substrate comprises approximately 5,000 discrete text fragments derived from the 1,232 papers and 4,753 compound records organized from the 33,269-compound library, although the exact number of QA/CoT pairs is not explicitly reported (Wang et al., 22 Jul 2025).
A defining architectural negative is as important as the positive description: the current system does not implement retrieval-augmented generation, document retrievers, ranking pipelines, or knowledge graphs (Wang et al., 22 Jul 2025). It is therefore best understood as a domain-adapted generative model trained on curated instruction data, not as a hybrid retrieval-and-reasoning system. This distinction matters because some apparent capabilities—such as literature synthesis or ranking—are learned through supervised domain adaptation and prompt structure rather than through explicit online retrieval (Wang et al., 22 Jul 2025).
4. Reasoning workflow and prompt-mediated experimental design
Perovskite-R1’s practical behavior is organized around a structured prompt strategy comprising three explicit components: Task Definition, Objective Criteria, and Expected Output Specification (Wang et al., 22 Jul 2025). The prompts can also include additional filtering rules such as “Prefer commercially available or natural/food-grade compounds,” “Solubility in DMF/DMSO,” and “If fewer than 3 candidates meet all criteria, do not force the output,” with final selections formatted inside braces containing chemical name and CAS number (Wang et al., 22 Jul 2025). This prompt discipline functions as a lightweight protocol language for translating mechanistic goals into reproducible model outputs.
At the chemical-reasoning level, the system maps functional groups to plausible additive mechanisms. Carboxylates, hydroxyls, pyridines, phosphine oxides, nitro groups, and amide or carbamate functionalities are connected to Lewis acid–base interactions with , ionic coordination to undercoordinated metal sites, hydrogen bonding relevant to halide-vacancy suppression, and transient coordination effects that modulate precursor-solution crystallization (Wang et al., 22 Jul 2025). The paper does not enumerate defect symbols such as or , but it explicitly frames the reasoning around defect passivation and crystallization kinetics in defect-relevant regimes (Wang et al., 22 Jul 2025).
The workflow is explicitly closed loop. The reported framework proceeds through four stages: curation of literature and candidate compounds with instruction/CoT generation; LoRA fine-tuning of QwQ-32B to obtain Perovskite-R1; structured-prompt screening and experimental protocol drafting; and laboratory synthesis and device fabrication followed by performance measurement and feedback (Wang et al., 22 Jul 2025). In that sense, the model is not merely an answer engine. It is part of a workflow for iterative hypothesis refinement in which AI-recommended additives can be compared against manually selected ones under identical experimental conditions (Wang et al., 22 Jul 2025).
The example outputs reflect this mechanism-aware screening style. In the Appendix, the model reasons through many candidates and ranks 5-hydroxy-2-methylbenzoic acid, dimethylphosphine oxide, and 2-methoxy-4-nitrophenol on mechanistic grounds (Wang et al., 22 Jul 2025). This does not imply automatic protocol completeness, however. A recurring theme in the paper is that Perovskite-R1 is strongest at mechanistic ranking and candidate selection, not at enforcing all synthesis constraints automatically (Wang et al., 22 Jul 2025).
5. Experimental validation and benchmark performance
The experimental validation uses a mixed-cation iodide perovskite, $32$0, with additive concentration fixed at $32$1 in the precursor solution (Wang et al., 22 Jul 2025). Two AI-proposed additives were tested—3,5-difluoropyridine-2-carboxylic acid (AI-DFCA) and 5-hydroxy-2-methylbenzoic acid (AI-HMBA)—against two manually chosen baselines, gallic acid (Manual-GA) and caffeic acid (Manual-CA) (Wang et al., 22 Jul 2025). JV measurements were performed under AM 1.5G illumination at $32$2 with scan rate $32$3 per $32$4, while the paper does not report specific ETL/HTL stacks, solvent identities, or annealing parameters for this validation set (Wang et al., 22 Jul 2025).
The control device gave $32$5, $32$6, $32$7, and $32$8 (Wang et al., 22 Jul 2025). Manual-GA degraded performance to $32$9, $0.1$0, $0.1$1, and $0.1$2, while Manual-CA yielded $0.1$3, $0.1$4, $0.1$5, and $0.1$6 (Wang et al., 22 Jul 2025). By contrast, AI-DFCA reached $0.1$7, $0.1$8, $0.1$9, and $1$0, and AI-HMBA reached $1$1, $1$2, $1$3, and $1$4 (Wang et al., 22 Jul 2025). The immediate implication is modest improvement relative to the control together with a clear advantage over manually chosen literature-inspired additives in this specific system and loading (Wang et al., 22 Jul 2025).
The benchmark study shows the same pattern at the QA level. Domain experts assessed a perovskite-focused benchmark split by difficulty, and Perovskite-R1 exceeded all compared general models across easy, medium, and hard subsets (Wang et al., 22 Jul 2025).
| Model | Accuracy (easy / medium / hard) |
|---|---|
| QwQ-32B | 76.27 / 76.84 / 74.75 |
| Doubao-seed-1.6-thinking | 71.30 / 71.46 / 70.77 |
| DeepSeek-R1 | 76.42 / 74.88 / 73.97 |
| Gemini-2.5-flash-thinking | 78.71 / 77.52 / 76.05 |
| Perovskite-R1 | 86.92 / 85.06 / 84.63 |
The paper states that the gain over the original QwQ-32B is about 10% absolute and is most pronounced on hard, domain-specific questions, which the authors attribute to curated instruction-tuning with chain-of-thought data (Wang et al., 22 Jul 2025). At the same time, several common validation desiderata are absent: statistical significance, sample sizes, hysteresis indices, trap-density measurements, stability metrics, ablations, and detailed failure analyses are not reported (Wang et al., 22 Jul 2025).
6. Interpretation, limitations, and projected extensions
Perovskite-R1 is significant chiefly as a proof that domain-adapted LLMs can couple literature synthesis to materially nontrivial laboratory decisions in PSC research (Wang et al., 22 Jul 2025). It should not, however, be conflated with an autonomous materials-discovery stack. The paper explicitly notes that output depth can be limited in single-turn interactions, that protocol parameterization still requires manual refinement, and that feasibility constraints are not automatically enforced (Wang et al., 22 Jul 2025). This places the model in a researcher-in-the-loop regime rather than a closed autonomous lab regime.
Several misconceptions are corrected by the reported results. One is that general-purpose reasoning models are sufficient for specialist perovskite tasks; the benchmark indicates otherwise (Wang et al., 22 Jul 2025). Another is that literature-precedented additives will necessarily transfer across perovskite compositions and processing windows; Manual-GA and Manual-CA degraded device performance under the tested conditions despite their broader literature familiarity (Wang et al., 22 Jul 2025). A third is that the present system already derives its competence from retrieval or graph tooling; in fact, those features are not yet implemented (Wang et al., 22 Jul 2025).
The limitations are equally explicit. Dataset coverage and literature bias remain inherent, there is no explicit bias correction or uncertainty quantification, and the risk of hallucination remains in open-ended reasoning (Wang et al., 22 Jul 2025). No ablation distinguishes the relative contribution of instruction tuning versus chain-of-thought, and no detailed error taxonomy is provided (Wang et al., 22 Jul 2025). Resource availability is stronger than methodological closure: the authors provide an example of the training set on Hugging Face, identify QwQ-32B as the base model, and report fine-tuning with LLaMA-Factory, but they do not report hardware or training duration (Wang et al., 22 Jul 2025).
The future-work agenda is correspondingly broad. Proposed extensions include multi-turn dialogue design and RLHF; knowledge-graph-based feasibility checking; integration of physics and simulation models; expansion to interface engineering, solvent optimization, stability regulation, and device-architecture co-design; and extension to low-dimensional perovskites, tandem architectures, and wide-bandgap stability (Wang et al., 22 Jul 2025). Integration with automated synthesis platforms and high-throughput experimentation is also proposed (Wang et al., 22 Jul 2025). This suggests that Perovskite-R1 is best viewed not as a finished expert system, but as an early domain-specific reasoning layer for a more tightly coupled materials-design loop in perovskite photovoltaics.