MXLLM: An integrated literature-to-experiment framework for MXene electromagnetic absorbers

The underutilization of unstructured experimental data in academic literature limits AI-assisted materials research, particularly for complex multicomponent systems such as MXene-based electromagnetic absorbers. Here, we present MXLLM, an integrated four-stage framework comprising (1) literature acquisition, (2) schema-guided extraction, (3) reflection-based quality control, and (4) threshold-based ML screening with source-linked protocol generation and experimental verification. To mitigate hallucination and numerical extraction errors, a Discriminator–Optimizer reflection loop audits and iteratively repairs schema-constrained JSON records, enabling construction of a database containing 3621 entries across 68 fields from 2153 papers with an extraction F1 score of 98.3%. The application stage combines semantic retrieval with two parallel XGBoost classifiers for threshold-based screening of novel compositions; under paper-level splitting, the reflection loss (RL) and effective absorption bandwidth (EAB) classifiers achieved AUC values of 0.908 and 0.887, respectively. The framework is demonstrated through retrospective retrieval of Ti₃C₂T_x/CoFe@C and prospective evaluation of a Ti₃C₂T_x/NiFe@C candidate that we proposed by chemical analogy, with EAB = 5.47 GHz. These cases demonstrate an integrated data-mining-to-verification workflow whose transferability beyond the present Ti₃C₂T_x systems requires further validation.