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…
Nature reported that an autonomous closed-loop framework combining active learning, quantum modelling, Bayesian optimization, and symbolic regression explored passivation molecules together with fabrication processes and identified 5ANI, achieving PCE 27.22% in 0.05 cm² cells and 23.49% in 21.4 cm² mini-modules.
OpenAI announced GPT-6 for everyone and introduced Intelligent UI in ChatGPT. The new UI is intended to support learning and task execution by presenting answers in a more visual and interactive form. OpenAI said the move broadly deploys a next-generation model to ChatGPT, which is used by more than 1.2 billion people per week.
Anthropic released Claude Haiku 5.5. The company describes it as its cheapest and fastest small model, suitable for high-frequency, low-cost work, rapid customer support, browser use, and sub-agent-style coding. Anthropic said the price was reduced by about 75% compared with Haiku 4.5.
A npj Computational Materials paper compared fine-tuning strategies such as targeted fine-tuning, transfer learning, multihead fine-tuning, LoRA, and L²-SP to predict phonon band structure, thermal and elastic properties, and the potential-energy surface of unstable phonon modes across 53 materials. Performance could be improved with only a small number of material-specific training structures, and some strategies improved thermal and elastic properties while reducing phonon error.
OpenAI unveiled GPT-6 Astra on 2026-09-29. It highlighted improved performance in computer use, browsing, software engineering, cybersecurity, and scientific tasks, and said it would first deploy to some organizations before expanding to ChatGPT paid and enterprise plans, API, Azure, and Bedrock.
Physics-guided active learning combining Bayesian optimization, XGBoost regression, and RFE was used to predict interface barriers from literature-based data. The approach narrowed candidates to 5 composite systems and led to experimental validation, with a test R² of 0.87 and, in PEDOT:PSS/EGaIn, a 0.24 eV barrier and an 8.80 mV open-circuit voltage at 25 K being reported.
Atom-JEPA was proposed as a joint-embedding predictive architecture for label-free representation learning of 3D atomic structures. After pretraining, the model was transferred to molecular ADMET, quantum chemistry, and crystal property prediction, and it was reported to achieve top performance in ADMET and quantum chemistry while remaining competitive on crystal property prediction.
OpenAI released the GPT-6 Astra model page, reflecting a 2026-09-29 update. The company emphasized strengths in computer use, browsing, software engineering, cybersecurity, science, and work-related tasks, and said deployment would start with some organizations and expand to Plus, Pro, Business, Enterprise, API, Azure, and Bedrock within days.
Mistral released the new model Mistral Large 4 on 2026-10-06. The company said it showed strengths in benchmarks related to cybersecurity, agent behavior, and finance and legal workflows, and stated that the public preview would be provided on its European data center infrastructure. It plans to release the weights later this month.
OpenAI publicly disclosed on 2026-10-05 how it will respond to EU text provenance rules. Around the same time, notices related to safety and governance were also posted, indicating a trend toward simultaneous strengthening of regulatory compliance and model deployment systems.
According to Nature Communications(2026-10-06), a diffusion model that directly incorporates physical constraints into the generative process produced 3D velocity fields of rotating turbulence. The method better reproduced anisotropic energy spectra and intermittency statistics while maintaining physical conditions such as incompressibility, and it converged faster in training than the existing DDPM.
Anthropic announced that it will invest $100 million and train 10,000 engineers to address the AI talent shortage faced by companies. The announcement focuses on workforce development and expanding enterprise adoption rather than launching a new model.
An ACS Applied Materials & Interfaces paper reported that, after extracting microstructural features from SEM images, it combined cluster analysis with supervised and unsupervised ML to classify perovskite solar cell candidates and derived high-performance design rules through experimental validation. It is a case that combines materials-informatics exploration with experimental confirmation.
A Physical Review B paper theoretically analyzed the problem of identifying topological properties in non-Hermitian (topological) systems using supervised-learning machine learning, showing that generalization can be limited in high-dimensional non-Hermitian situations because of the nonlinear abrupt change of non-Bloch topological invariants. It presented a valid prediction range through generalization error bounds, but no direct validation with experimental materials data was reported.
A study was reported that directly maps antiferroelectric domain evolution under an electric field and predicts equilibrium microstructures using a scale-invariant physics-informed neural operator. It accelerated large-scale computation and showed zero-shot generalization, but validation is needed outside the physical constraints and training range.