Artificial intelligence is moving beyond theoretical benchmarks and into industrial discovery, with companies now using AI to identify new materials in hours that would traditionally take months or years to find. At the International Conference on Machine Learning (ICML) 2026 in Seoul, LG AI Research showcased how its proprietary AI model, Exaone, discovered a novel hair loss management compound called Rhamsydil from over 420,000 candidate materials in just one day, marking a significant shift in how AI is being applied to real-world manufacturing and product development.

What Makes This Different From Previous AI Breakthroughs?

The key distinction here is that Exaone Discovery isn't simply generating theoretical predictions; it's operating within a closed-loop system that connects AI screening to experimental validation. The platform uses machine learning to autonomously recognize molecular structures from scientific literature and design candidate substances with desired performance characteristics. LG completed patent registration for this entire R&D process earlier this year, establishing what the company calls a technological barrier.

This approach mirrors a parallel discovery workflow reported by researchers at Aalto University, who used machine-learning-guided screening combined with first-principles calculations to identify two new kagome superconductors, YRu3B2 and LuRu3B2. The critical insight from both efforts is that AI's value isn't in replacing physics or experiments; it's in dramatically narrowing the search space so that expensive computational validation and synthesis can be targeted at the most promising candidates.

"Exaone is no longer a lab-based AI but an expert AI that finds answers in the field," stated Lim Woo-hyung, Co-Director of LG AI Research.

Lim Woo-hyung, Co-Director of LG AI Research

How Are Companies Implementing AI-Driven Materials Discovery?

• Multi-Stage Screening Process: LG's Exaone Discovery uses AI to prescreens hundreds of thousands of molecular candidates, then passes only the highest-ranked options to more expensive first-principles calculations and experimental synthesis, reducing the number of compounds that need physical testing.
• Physics-Informed Feature Engineering: Rather than relying on generic chemical similarity alone, the AI systems encode domain-specific knowledge about molecular behavior, such as electron-phonon coupling and flat-band effects, to improve ranking precision at the top of candidate lists.
• Closed-Loop Validation: The workflow connects AI predictions directly to laboratory synthesis and measurement, ensuring that AI-generated candidates can be experimentally confirmed before commercialization begins.

The Rhamsydil discovery exemplifies this workflow. LG Household & Health Care collaborated with LG AI Research to identify a hair loss prevention compound that lacks steroid-derived ingredients, a property that would have been difficult to specify using traditional high-throughput screening alone. The company has already presented this achievement at the World Congress for Hair Research and is preparing for product commercialization.

Beyond materials discovery, LG demonstrated the breadth of Exaone's industrial applications. The company co-developed an immersion cooling fluid material with GS Caltex to address heat generation in AI data centers, and launched Exaone BI, a finance-specialized AI agent that analyzes approximately 8,000 stocks daily and recently signed a contract with Koscom to provide domestic South Korean stock prediction services.

Why Does the Speed of Discovery Matter?

The economics of materials research have traditionally been constrained by the cost and time required for experimental validation. By using AI to reduce the number of candidates that need expensive synthesis and testing, companies can explore larger chemical search spaces without proportionally increasing R&D budgets. LG's Exaone Data Foundry platform, which generates high-quality training data and automatically builds specialized AI models, has achieved at least a 1,000-fold increase in data productivity and an average quality improvement of over 20 percent.

The superconductor research from Aalto University reinforces this point. While the identified compounds, YRu3B2 and LuRu3B2, operate only at cryogenic temperatures rather than room temperature, the demonstrated workflow shows how AI can accelerate the discovery process itself. The team used ML to narrow the search space, then applied targeted quantum-geometry calculations before Rice University collaborators synthesized and tested the candidates. For materials-discovery practitioners, the lesson is that physics-informed descriptors and top-ranked precision matter more than generic benchmark accuracy when AI systems must produce lab-verifiable discoveries.

LG AI Research has published 363 papers at top-tier academic conferences since its inception in December 2020 and filed 838 patents, including 371 in South Korea, 243 overseas, and 224 international applications. The institute presented 14 papers at ICML 2026, with its new material generative AI technology ranking second globally in the LeMat-GenBench comprehensive benchmark, which evaluates the stability of AI-generated crystal structures, their distinctiveness from existing materials, and the diversity of proposed candidate substances.

The shift toward AI-accelerated materials discovery signals a broader trend in industrial AI: moving away from general-purpose benchmarks and toward domain-specific applications that solve concrete business problems. As Lim Woo-hyung noted, the goal is to ensure that research outcomes do not remain confined to academic papers but are applied to actual industrial sites, where they can generate measurable economic value.