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…
Google has unveiled Gemini 4 Argon. The announcement highlighted a 1,000,000-token context window, a design for multi-step reasoning problem solving, and a priority on broader pre-release safety validation. The emphasis on both large-scale context and validation procedures stood out.
A case listed on Nature’s materials ML topic page as a 2026-09-17 public entry introduces a 2D material manufacturing approach that connects AI to CVD synthesis, quality control, and autonomous manufacturing. It is classified as a real-world process application, but the publicly available information is limited to the topic-page level, so detailed performance metrics cannot be confirmed.
Anthropic announced an independent evaluation collaboration with Accenture's Faculty on 2026-09-18 to conduct model evaluation, red teaming, alignment evaluation, and safety guardrail testing. The two sides said they plan to invest at least 1 billion dollars each over the next 5 years.
WeedNet, published in Nature Communications, is a foundation model that combines self-supervised learning, fine-tuning, and reliability-improvement strategies in a global-local hybrid pipeline to identify and classify images of 1,593 weed species. The reported overall accuracy was 91.02%, and it reached 97.38% for 84 species in the Iowa region. However, this is a weed image classification case and is not a materials-science application.
Anthropic released Claude Sonnet 5.5 on 2026-09-28. The company said it is 30% faster than Sonnet 5 and up to 30% cheaper for most tasks. This shows that model performance competition is expanding to include speed and cost efficiency.
OpenAI introduced GPT-6.1 Sol on 2026-09-29. DevDay 2026 Recap was also posted on the same day, but the key point of this latest news item is the new model release. This shows that OpenAI’s flagship LLM lineup continues to be refreshed.
Google DeepMind’s September 2026 news list newly includes Gemini 3.8 Flash, Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, and improvements to Private AI Compute including secure server-side memory. Based on the list page, only the fact that an official announcement was made in September can be confirmed.
Anthropic released Claude Opus 5.5 on 2026-09-22. The company explained that agentic coding, computer use, and knowledge work have improved, and that automated behavior auditing performance has also become stronger. Biological and cyber capabilities are provided only in a limited way together with a verification program, and the price has also been lowered compared with the previous Opus 5.
A Neural ODE based on neural networks directly generated control pulses and designed and experimentally implemented a π/2 gate in a superconducting transmon circuit. The reported performance was about 99.95% gate fidelity and detuning robust performance at the ±20 MHz level.
Researchers performed microstructure inverse design aimed at a strength-ductility synergy by combining physics-aware generative optimization, which incorporates interfacial and crystallographic information into the loss, with latent-space sampling. The generated structures included combinations outside the training data, and they were validated through crystal plasticity finite element simulation and experiments.
A Nature Communications paper used MALVINA, a deep learning-based single-cell profiling workflow, to measure bacterial invasion and host DNA damage simultaneously at the single-cell level and to classify and compare pathogenicity changes. It reported identifying strain-to-strain interactions and virulence remodeling under drug treatment.
An npj Computational Materials paper reported a machine-learning workflow that predicts and interprets spectroscopic properties such as absorption, emission, and lifetime from local coordination structures in Nd/Y co-doped fluorites. Linear regression, SVR, BPNN, XGBoost, and SHAP were used to quantify structure-property correlations.
Anthropic said that it used Claude to discover a new enzyme system that shows CRISPR-like properties. This is an example showing that AI can contribute in practice to life-science exploration and hypothesis generation beyond simple question answering.
OpenAI disclosed a model misalignment reporting framework on 2026-09-16 and presented 6 anomalous behavior cases from the past 6 months. The key point is that it has defined an operational principle of systematically disclosing abnormal behavior.
Google DeepMind disclosed on 2026-09-23 that it has introduced server-side memory to Private AI Compute. The structure aims to keep long-term context and memory in the cloud while emphasizing privacy on the premise of device-side keys, end-to-end encryption, and secure execution environments.