Machine Learning Identifies an Information-Sufficiency Regime for Critical-Mineral Recovery
Abstract
The recovery of critical minerals from spent lithium-ion batteries is governed by coupled feedstock chemistry, solvent composition, and processing conditions, yet heterogeneous experimental reporting makes it difficult to determine which information is essential for reliable prediction across heterogeneous experimental systems. Here, using deep eutectic solvent-based recovery as a representative resource recovery platform, we curate 814 experimental observations from 38 publications and develop an interpretable machine learning framework to identify an informationsufficiency regime. Five descriptors, including temperature, electrode material type, leaching time, target metal, and hydrogen-bond donor identity, retain 98.3% of the predictive performance of the full 13-descriptor model while reducing the descriptor space by 61.5%. Reducing the representation from five to four descriptors causes a pronounced performance loss, whereas adding a sixth provides negligible additional benefit. Descriptor removal and domain ablation analyses further show that temperature and target metal identity contain the least replaceable predictive information, while material identity, solvent chemistry, and processing conditions provide complementary information. These results identify a compact, chemically structured information core for critical-mineral recovery and provide a quantitative basis for prioritizing experimental measurements, improving cross-study data consistency, and developing more transferable datadriven frameworks for battery resource recovery.
Related
- Conventional Ion-Exchange and Chelating Resins for Metal Recovery: A Critical Comparison of Selectivity, Dynamic Performance, Regeneration, and Process Integration
- From Agri-Food Byproducts to High-Value Bioactive Compounds: A Critical Review Linking Green Recovery and Chemical Profiling to Circular Valorization
- Surfactants for enhanced oil recovery: A critical review of mechanisms, performance, reservoir compatibility, and field applications
- Deep learning identifies transversely isotropic material properties using kinematics fields
- Epoxy-Based Ion-Exchange Materials for Water Treatment and Resource Recovery: A Critical Review of Structure–Performance Relationships and Practical Limitations
BibTeX
@misc{liu2026machine,
title = {{Machine Learning Identifies an Information-Sufficiency Regime for Critical-Mineral Recovery}},
author = {Tengxiao Liu and Ta-Wei Wang and Hongtao Sun},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010566/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010566/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Machine Learning Identifies an Information-Sufficiency Regime for Critical-Mineral Recovery
AU - Tengxiao Liu
AU - Ta-Wei Wang
AU - Hongtao Sun
PY - 2026
DA - 2026/10/16
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010566/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010566/v1
AB - The recovery of critical minerals from spent lithium-ion batteries is governed by coupled feedstock chemistry, solvent composition, and processing conditions, yet heterogeneous experimental reporting makes it difficult to determine which information is essential for reliable prediction across heterogeneous experimental systems. Here, using deep eutectic solvent-based recovery as a representative resource recovery platform, we curate 814 experimental observations from 38 publications and develop an interpretable machine learning framework to identify an informationsufficiency regime. Five descriptors, including temperature, electrode material type, leaching time, target metal, and hydrogen-bond donor identity, retain 98.3% of the predictive performance of the full 13-descriptor model while reducing the descriptor space by 61.5%. Reducing the representation from five to four descriptors causes a pronounced performance loss, whereas adding a sixth provides negligible additional benefit. Descriptor removal and domain ablation analyses further show that temperature and target metal identity contain the least replaceable predictive information, while material identity, solvent chemistry, and processing conditions provide complementary information. These results identify a compact, chemically structured information core for critical-mineral recovery and provide a quantitative basis for prioritizing experimental measurements, improving cross-study data consistency, and developing more transferable datadriven frameworks for battery resource recovery.
ER -