MSP-LLM: A Structured Large Language Model Framework for Inorganic Material Synthesis Planning
Organizations: Department of Industrial & Systems Engineering, KAIST, Daejeon, Korea · Graduate School of Data Science, KAIST, Daejeon, Republic of Korea · Korea Research Institute of Chemical Technology, KRICT, Daejeon, Korea
Abstract
Material synthesis planning (MSP) remains a fundamental and underexplored bottleneck in AI-driven materials discovery, as it requires not only identifying suitable precursor materials but also designing coherent sequences of synthesis operations to realize a target material. Although several AI-based approaches have been proposed to address isolated subtasks of MSP, a methodology for solving the entire MSP task has yet to be established. We propose MSP-LLM, a structured LLM-based framework that formulates MSP as a two-stage process composed of two constituent subproblems: precursor prediction (PP) and synthesis operation prediction (SOP). Our approach introduces a discrete material class as an intermediate decision variable that organizes both tasks into a chemically consistent decision chain. For SOP, we further incorporate hierarchical precursor types as synthesis-relevant inductive biases and employ precursor constraint factorization that preserves precursor-related information in the autoregressive decoding state. Extensive experiments show that MSP-LLM consistently outperforms existing methods on both PP and SOP, as well as on the MSP task, demonstrating an effective and scalable framework with practical potential for autonomous MSP.
Figures & tables
| Method | Acc. | |
| Target only | 0.6502 | — |
| Implicit | 0.6838 | +3.4 pp |
| Explicit (PCF) | 0.8479 | +16.4 pp |
| (a) Precursor Prediction (PP) | ||||||
| Model | Top-1 | Top-3 | Top-5 | Top-10 | ||
| (A) Naive Fine-tuning | 70.59 | 84.90 | 90.62 | 94.39 | ||
| (B) + Material Group ( MSP-LLM ) | ✓ | 70.75 | 85.56 | 91.68 | 94.97 | |
| (b) Synthesis Operation Prediction (SOP) | ||||||
| Model | EC | Top-10 | NED | LCS | F-1 | |
| (A) Naive Fine-tuning | 11.24 | 0.5336 | 0.6311 | 0.6572 | ||
| # | Op. | (P/R, °C) | (P/R, h) | # | Op. | (P/R, °C) | (P/R, h) |
| Case 1: Ba 0.774 Ca 0.226 Ti 0.9 Sn 0.1 O 3 | Case 2: Nd 0.7 Sr 0.3 Mn 0.9 Ga 0.1 O 3 | ||||||
| Pre.: BaCO 3 , CaCO 3 , SnO 2 , TiO 2 | Pre.: Ga 2 O 3 , MnO 2 , Nd 2 O 3 , SrCO 3 | ||||||
| Ops: mix dry heat mix shape sinter | Ops: mix heat mix heat shape sinter | ||||||
| 0 | heat | 1108 / 1200 | 3.9 / 4.0 | 0 | heat | 1066 / 1000 | 24.3 / 24.0 |
| 1 | sinter | 1347 / 1400 | 3.4 / 4.0 | 1 | heat | 1173 / 1350 | 24.7 / 24.0 |
| 2 | sinter | 1398 / 1350 | 22.4 / 24.0 | ||||
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| LLM | Task | Split 1 | Split 2 | |||||
| LoRA Rank ( ) | LoRA Alpha | lr / bs | LoRA Rank ( ) | LoRA Alpha | lr / bs | |||
| LLaMA-3.1-8B | ||||||||
| [ 10 ] | PP | 16 | 32 | 1e-4 / 4 | 16 | 32 | 1e-4 / 2 | |
| LLaMA-3.1-8B | ||||||||
| [ 10 ] | SOP | 16 | 32 | 1e-4 / 2 | 16 | 32 | 1e-4 / 2 | |
| Qwen-2.5-7B | ||||||||
| Context | Qwen-2.5-7B | LLaMA-3.1-8B | |||||||
| Train | Test | (a) Synthesis Operation Prediction (SOP) | |||||||
| Top-10 | NED | LCS | F-1 | Top-10 | NED | LCS | F-1 | ||
| No | No | 29.94 | .732 | .816 | .843 | 29.38 | .728 | .816 | .844 |
| Yes | No | 29.66 | .730 | .815 | .839 | 28.92 | .726 | .812 | .836 |
| Yes | Yes | 32.18 | .742 | .824 | .849 | 31.44 | .742 | .824 | .850 |
| (b) Material Synthesis Planning (MSP) | |||||||||
| PP | SOP | MSP | ||||||||||||
| Method | Top-1 | Top-3 | Top-5 | Top-10 | Top-10 | NED | LCS | F-1 | Top-1 | Top-3 | Top-5 | Top-10 | ||
| Large Language Models (API, few-shot prompting) | ||||||||||||||
| GPT-4o-mini (20-shot) | 54.85 | 63.40 | 67.91 | 70.34 | 7.09 | 0.5847 | 0.7089 | 0.7407 | 0.19 | 0.28 | 0.28 | 0.47 | ||
| GPT-4o-mini (40-shot) | 55.97 | 64.27 | 67.82 | 71.36 | 7.93 | 0.5934 | 0.7141 | 0.7442 | 0.27 | 0.27 | 0.37 | 0.64 | ||
| GPT-4o-mini (60-shot) | 55.69 | 65.02 | 68.75 | 72.20 | 10.54 | 0.6099 | 0.7310 | 0.7616 | 1.40 | 1.96 | 2.05 | 2.24 | ||
| GPT-4o-mini (80-shot) | 59.51 | 67.44 | 71.92 | 75.37 | 9.24 | 0.6048 | 0.7238 | 0.7551 | 0.93 | 1.21 | 1.31 | 1.40 | ||
| Method | Top-1 | Top-3 | Top-5 | Top-10 |
| Naive two-stage fine-tuning | 4.1 | 8.3 | 9.9 | 12.7 |
| End-to-end fine-tuning | 4.8 | 10.8 | 13.0 | 16.1 |
| Nearest-neighbor | 5.8 | 9.8 | 10.3 | 10.8 |
| MSP-LLM | 6.72 | 12.12 | 16.60 | 23.23 |
| Element coverage | Mass balance | |||||||
| Model | @1 | @3 | @5 | @10 | @1 | @3 | @5 | @10 |
| MSP-LLM (Qwen-2.5-7B) | 99.6 | 99.9 | 100.0 | 100.0 | 99.4 | 99.8 | 99.9 | 100.0 |
| MSP-LLM (LLaMA-3.1-8B) | 99.6 | 100.0 | 100.0 | 100.0 | 99.3 | 99.8 | 99.8 | 100.0 |
| PP Criterion | SOP Condition | Top-1 | Top-3 | Top-5 | Top-10 |
| PP Set-F1 Criterion | |||||
| 0.8 (75.0%) | SOP Exact | 6.90 | 12.50 | 17.07 | 23.97 |
| 0.8 | SOP NED 0.8 | 12.31 | 20.24 | 25.00 | 32.74 |
| 0.8 | SOP LCS 0.8 | 17.82 | 30.88 | 38.06 | 49.35 |
| 0.8 | SOP F-1 0.8 | 21.36 | 36.01 | 43.47 | 55.60 |
| PP Elem-F1 Criterion | |||||
| Class | n | PP Top-10 | SOP Top-10 | MSP Top-10 (strict) | MSP Top-10 (NED 0.8) | MSP Top-10 (LCS 0.8) | MSP Top-10 (F1 0.8) |
| oxide | 938 | 95.3 | 25.59 | 18.0 | 29.00 | 45.52 | 51.28 |
| alloy | 39 | 100.0 | 38.46 | 35.9 | 43.59 | 53.85 | 58.97 |
| phosphate | 28 | 100.0 | 21.43 | 3.6 | 7.14 | 32.14 | 35.71 |
| pyrophosphate | 14 | 92.9 | 28.57 | 14.3 | 14.29 | 21.43 | 21.43 |
| oxyhalide | 9 | 66.7 | 44.44 | 11.1 | 11.11 | 22.22 | 22.22 |
| selenide | 10 | 100.0 | 10.00 | 10.0 | 10.00 | 80.00 | 90.00 |
| Class | n | PP Top-10 | SOP Top-10 | MSP Top-10 (strict) | MSP Top-10 (NED 0.8) | MSP Top-10 (LCS 0.8) | MSP Top-10 (F1 0.8) |
| oxide | 907 | 96.9 | 30.10 | 21.9 | 31.75 | 48.18 | 54.47 |
| alloy | 61 | 100.0 | 45.90 | 42.6 | 42.62 | 52.46 | 57.38 |
| phosphate | 28 | 85.7 | 21.43 | 3.6 | 7.14 | 21.43 | 28.57 |
| pyrophosphate | 9 | 88.9 | 55.56 | 22.2 | 22.22 | 44.44 | 44.44 |
| oxyhalide | 11 | 81.8 | 18.18 | 27.3 | 36.36 | 45.45 | 45.45 |
| selenide | 7 | 100.0 | 42.86 | 42.9 | 42.86 | 57.14 | 71.43 |
| Task | Split 1 | Split 2 | |||
| Accuracy | F-1 | Accuracy | F-1 | ||
| PP | 96.0 | 0.83 | 96.9 | 0.80 | |
| SOP | 95.4 | 0.72 | 96.4 | 0.78 | |
| # | Op. | (P/R, °C) | (P/R, h) |
| Case 3: Mg 0.94 Ni 0.06 ZrNb 2 O 8 (DOI: 10.1007/s10854-017-7996-x) | |||
| Pre.: MgO, Nb 2 O 5 , NiO, ZrO 2 | |||
| Ops: mix dry heat mix dry mix shape sinter | |||
| 0 | heat | 1026 / 1100 | 4.5 / 2.0 |
| 1 | sinter | 1351 / 1320 | 4.9 / 4.0 |
| Case 4: Ni 0.78 Cu 0.20 Zn 0.02 Fe 1.96 O 3.94 (DOI: 10.1016/j.jmmm.2017.12.108) | |||
| (a) Precursor Prediction (PP) | ||||
| Model | Top-1 | Top-3 | Top-5 | Top-10 |
| MSP-LLM (Qwen-2.5-7B) | ||||
| MSP-LLM (LLaMA-3.1-8B) | ||||
| (b) Synthesis Operation Prediction (SOP) | ||||
| Model | Top-10 | NED | LCS | F-1 |
| MSP-LLM (Qwen-2.5-7B) | ||||
| Model | Heat ops (total) | abs | rel | changed | among changed (%) |
| Implicit | 945 909 | 140 | 31 / 55 / 14 | ||
| Explicit (PCF) | 689 620 | 128 | 52 / 39 / 9 |
| Method | Split 1 | Split 2 | |||||||
| Top-1 | Top-3 | Top-5 | Top-10 | Top-1 | Top-3 | Top-5 | Top-10 | ||
| Specialized Models | |||||||||
| Retrieval based He et al. [13] | 1.46 | 3.75 | 5.94 | 10.42 | 1.12 | 3.54 | 5.50 | 11.19 | |
| SPENDE [ 23 ] | 3.66 | N/A | N/A | N/A | 5.71 | N/A | N/A | N/A | |
| Large Language Models (API) | |||||||||
| GPT-4o-mini (20-shot) | 1.73 | 3.11 | 4.2 | 5.85 | 1.68 | 3.26 | 4.01 | 7.09 | |
Materials Science
Materials science studies how composition, structure, processing, and defects determine material properties and how to design useful materials.
See all 330 papers