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…
A Nature Communications paper reported a method that uses a visual large language model to compare and rank candidate domain segmentation results for spatial transcriptomics, then combines the top results to create a consensus spatial domain map. It claimed results closer to expert annotation, more reproducible rankings, and better interpretability.
OpenAI expanded its security bug bounty program on 2026년 11월 2일. It broadly increased rewards for vulnerability reports, introduced additional incentives for high-quality reports and long-term contributors, and strengthened reward caps for critical vulnerabilities in core products. The move is a security measure aimed at attracting more external researchers.
A paper accepted in Physical Review B reported a deep learning method that predicts low-temperature conductance and topological visibility using nanowire conductance data measured at high temperature. A diffusion model based on a Shifted Window U-Net Transformer reconstructs low-temperature conductance, and a Video Vision Transformer directly predicts topological visibility, demonstrating screening potential.
OpenAI launched a national Grid Protectors Program that adjusts data center power demand according to grid conditions. The key point is that new computing campuses will be designed to contribute to grid reliability and electricity price stability.
Researchers applied a machine learning model trained only on experimental data to quantum diamond magnetometer measurements and tracked magnetic signals caused by elevator motion in real time. The reported performance was training within 40 minutes, 10 Hz tracking, more than 80% accuracy, and 30 cm error tolerance. This case can be viewed as an example of directly applying ML to a time-series tracking problem in measurement data.
The accepted paper in APS Physical Review Materials systematically calculated the tensile behavior of graphene oxide as a function of oxidation degree and OH/O ratio using a MACE-based machine-learning interatomic potential fine-tuned with DFT data. It reduced the overestimation seen in ReaxFF and reproduced the decreasing trends in Young’s modulus and ultimate tensile strength more accurately, while showing that dehydration is important to fracture behavior under highly oxidized and hydrophilic conditions.
OpenAI announced on 2026-10-27 that it is expanding its partnership with Oracle to connect Oracle applications and data to ChatGPT, and to embed OpenAI models and tools into Oracle products and services. In an enterprise context, it highlighted integrations with ChatGPT Work and Codex, and also introduced Oracle AI Workspace, AI Security Mesh, AI SMB Bundle, and App Composer.
OpenAI said it has broadened the scope of frontier model misalignment monitoring across training, evaluation, internal use, and external deployment. It explained that, prompted by gaps identified in an internal cybersecurity evaluation, it introduced detection layers that use signals from conversation, reasoning, and tool use.
Google DeepMind unveiled Genie 4 on 2026년 10월 27일. The model is a world simulation model that emphasizes real-time interactive simulation, photorealism, fluid motion, and native audio, and is currently available as a research preview for trusted testers. It is planned to expand to Google AI Ultra, Gemini API, and Google AI Studio in the future.
OpenAI released the updated Model Spec on 2026-10-29 and redefined the standards for model behavior. The revised version was reorganized around five values: fidelity to human intent, minimizing harm, honesty and transparency, human control, and care.
An equivariant graph neural network called EdenGNN accelerated PAW-level electronic-structure calculations by predicting charge density and augmentation occupancies. A general-purpose model trained on Materials Project also predicted band structures of unknown materials well, demonstrating the potential for large-scale screening.
The Nature Communications paper used an AI-based platform called DeepDelivery to screen human proteins at large scale for scaffold candidates for mRNA delivery. Proteome-scale screening proposed 512 candidates, 15 were selected, and 12 acted as actual scaffold components. Interpretive analysis identified TRIM domain features, and TRIM25 domain engineering improved delivery performance.
Researchers at the MIT Energy Initiative presented a multimodal model that aligns textual descriptions of crystal structures with 3D structures using contrastive learning. The model showed that it can help with structural understanding and property inference for crystalline materials. However, this is an article-level public description, and reproducibility and generalizability in real materials development settings still require further validation.
A JACS paper presented a method for efficiently calculating Li electrode potentials in battery electrolytes by fine-tuning a general-purpose machine learning potential. This is an example of accelerating costly atomistic simulation tasks in materials calculations, such as potential evaluation. However, what the source showed was improved computational efficiency and applicability, not universal performance validation across a wide range of electrolytes.
An ACS Materials Letters paper combined nonadiabatic molecular dynamics and time-domain DFT results with interpretable machine learning to explain how regioisomeric halogenation changes the excited-state charge dynamics of DJ-phase halide perovskites. It derived the conclusion that substitution patterns are important for controlling charge dynamics.