AI & Materials
Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.Cases where AI was actually used in materials research and scientific experiments.
매일 00·06·12·18시(한국시간)에 웹에서 새 소식을 찾습니다. AI 활용 방법과 결과가 확인된 새 사례만 게시합니다.
Checking notifications…
This materials informatics study predicts piezoelectric tensors through learning that incorporates topological information and E(3)-equivariance. It demonstrates a direction for improving prediction consistency by embedding geometric symmetry and physical constraints into the model.
This materials informatics study evaluates candidate polymer dielectrics with interpretable screening at the substructure level and organizes the results into rational design rules. Because the approach emphasizes interpretability while narrowing the candidate pool, experimental synthesis and measurement must follow for verification.
A Chemistry of Materials paper published on 2026-09-28 addressed an approach that uses machine-learning interatomic potentials (MLIP) for solid-state batteries to model interfaces and structures at the atomic scale and accelerate related atomistic simulations. This method suggests a useful direction for exploring ion transport, interface evolution, and degradation mechanisms in solid-state battery interfaces with larger scale and higher accuracy.
A pipeline for exploring and discovering antiferromagnetic metal-organic framework (MOF) candidates using interpretable machine learning was reported. This is a materials informatics case that applies materials data to search for magnetic functional MOFs by jointly handling property signals and structural cues.
Ultrafast carrier dynamics under polarization control in sliding ferroelectrics were analyzed with ML-accelerated nonadiabatic dynamics simulations. This study is an example of using machine learning to accelerate nonadiabatic dynamics calculations and interpret how changes in polarization state affect carrier behavior.
SoLiD26 is a first-principles dataset for solid-liquid interfaces containing 15,400,000 structures, and it was shown to be applicable in training, validation, and test splits with a MACE-family MLIP. The key point is that it provides large-scale reference data for developing interatomic potentials for interface phenomena. However, performance outside the dataset distribution and split scheme requires separate confirmation.
A public benchmark called PolyBench26 was presented and compared polymer property prediction models using about 250,000 data points. Comparing language models, graph-based models, and descriptor-based models, the graph-based model showed the lowest error and robustness to increasing repeat-unit complexity. Because the result is benchmark-based, the ranking may change if the distribution of real research data differs.
It was reported that applying the SHRAV(State-Hypothesis-Reason-Action-Verify) framework to physical modeling and inverse design, especially computational lithography inverse design, increased thresholded aerial-image IoU from 0.5313 to 0.8153. The maximum absolute error between prediction and reproduction was also reported to be about 0.000824. However, because this is a single-case result, further validation is needed before generalizing it to other materials or device problems.
In a benchmark comparing 12 generative CSP models, template retrieval was strongest with a top-1 of 68.3%, and many generative models showed overlapping performance with template-substitution approaches. Removing a specific stoichiometric prototype family caused accuracy to drop by 50–78%, revealing that current generative crystal structure prediction is substantially prototype-dependent.
An agentic workflow converted user-defined interatomic potential equations into a LAMMPS pair style and carried out build, test, correction, and validation. It matched single-point energies and forces across 7 potential/material pairs, reproduced lattice constants, vacancy formation energies, and elastic constants on 5 crystalline benchmarks, and also reported nanosecond-scale simulation stability.
A study has been reported that simulated heat transfer at the Al/Al₂O₃ interface with non-equilibrium molecular dynamics using a neuroevolution potential trained from first-principles calculations, and interpreted inelastic phonon scattering through temperature-dependent interfacial thermal conductivity and spectral analysis. It agreed well with experimentally estimated values at low temperature and provided evidence for inelastic scattering.
Interpretable machine learning and SHAP were used to analyze the relationship between structural descriptors and negative thermal expansion properties in framework materials. A trade-off between large negative thermal expansion and a wide temperature range was identified, yielding structure-property trends that can guide design.
A method was reported for more efficiently simulating optical response based on the semiconductor Bloch equations by using a learned single-particle Hamiltonian in a non-resonantly photoexcited crystal. It was stated that calculation speed was greatly increased while maintaining the accuracy of current density and carrier occupation.
A method combining unsupervised learning and supervised learning was proposed to find topological branch points and transition locations in Majorana nanowires. Using unlabeled simulation data, the study showed the possibility of distinguishing topological and non-topological Majorana splitting and of building a classification tool for interpreting real experiments.
OMC-bench and AtomBit-OMC were released, targeting structure relaxation, cocrystal reconstruction, polymorph energy ranking, and OOD force/stress prediction in organic molecular crystals (OMC). It was reported that the cocrystal structure reconstruction failure rate was the lowest and that polymorph ranking performance was the best, and that important covalent and noncovalent contacts were interpreted with a weight map.