Powered by the upgraded RSI engine, the platform unifies data, models, experiments and production to accelerate materials innovation

BEIJING, Oct. 19, 2026 /PRNewswire/ -- Fabarta Co., Ltd., a leading provider of enterprise-grade AI agents and knowledge engines, today announced a major strategic upgrade to its AI for Science (AI4S) platform: the official launch of Fabarta Science Intelligent—the platform's newly upgraded core engine, driven by RSI (Recursive Self-Improvement)—alongside the evolution of AI4S into AI4S&E (AI for Science & Engineering), which now covers the core chain of R&D and production. The upgrade marks a comprehensive enhancement of Fabarta's AI4S platform, equipping it with frontier capabilities including multi-agent collaboration, AutoPINO model self-evolution and physics-informed AI hybrid modeling. The platform now connects the entire chain—from literature review and formulation design to dry- and wet-lab experimental validation, digital pilot-scale scale-up and industrial mass production—with the ability to iterate continuously, accelerating the shift of materials R&D from traditional trial-and-error empiricism to a new paradigm driven by AI and knowledge.

From Single-Point Tools to Full-Chain Integration: The AI4S Platform Takes a Leap with RSI

As an AI leader deeply engaged in the fine chemicals and new materials sectors, Fabarta had already achieved notable results with its earlier AI4S platform in scenarios such as intensive literature reading, data management and single-point algorithm-based prediction. Yet in the "deep-water zone" of the materials industry—moving from the laboratory to the production line—conventional AI4S tools often remain at the level of "single-point computing" or "assisting with paper publication."

To address this bottleneck, Fabarta systematically iterated on its AI4S platform, deeply embedding its self-developed RSI intelligent architecture into the platform foundation. Through its "Four Modules" and "Three Loops," the upgraded AI4S platform is now capable not only of literature Q&A and molecular generation, but also of adaptive model evolution, multiscale physical simulation and digital-twin calibration—marking the leap of Fabarta's AI4S platform into an "industry-grade, integrated R&D-production platform with self-iteration capability."

Breaking Two Long-Standing Bottlenecks in Materials R&D: Multi-Objective Balancing and Data Gap

New materials R&D has long been constrained by two bottlenecks: balancing competing multi-objective performance targets—such as the trade-offs among heat resistance, impact toughness, processability and cost—and the severe fragmentation of R&D and production data. Heat-resistant ABS resin is a case in point: introducing heat-resistant monomers usually comes at the expense of impact strength, while increasing rubber content restores toughness but lowers heat resistance. Meanwhile, critical clues scattered through historical literature, patents and "anomalous" batches remain siloed across departments and systems, making it difficult to establish a continuously operating "question–verify–adjust" loop.

"Fabarta has always been committed to building the industrial infrastructure of AI4S. The core of this major upgrade to our AI4S platform is the introduction of the RSI intelligent engine, which deeply integrates large models with our proprietary knowledge engine. Integrating AI into industrial upgrading cannot rely on a single general-purpose large model alone—AI must be embedded in the physical mechanisms and knowledge chains of the industry. We hope to help research and industrial teams worldwide, through the upgraded AI4S platform, connect the intelligent full chain spanning dry- and wet-lab bench-scale experiments, base pilot-scale trials and enterprise mass production, so that AI can create measurable industrial value for materials innovation." — Andy Gao, Founder and CEO of Fabarta

RSI's Core Operating Architecture: Four Modules Powering a Three-Layer Iterative Loop

Within the RSI platform, every materials research project is carried by four tightly integrated modules—Project, Session, Experiment and Resource Center—and driven by three layers of deep iterative loops that help R&D cross the barriers of the physical world:

• P0 Validation Loop (precise hypothesis validation, real-time model evolution): Within a single project session, the research agent draws on historical formulations and predictive models consolidated in the knowledge graph to rapidly generate testable hypotheses. The platform's built-in AutoPINO self-evolving model training engine, centered on physics-constrained neural operators and integrating geometry-adaptive perception, physics-residual-driven optimization and online incremental learning, maps molecular simulation, mesoscale phase-field simulation and macroscopic mechanical testing data into a unified multiscale physical representation space, progressively consolidating a composable, transferable and evolving library of domain neural operators. Whenever new experimental data is written back to the knowledge graph, the system dynamically identifies cognitive blind spots through graph-aware physical-consistency assessment and predictive-uncertainty analysis, adaptively triggering local operator updates, targeted retraining and model calibration, and feeds the updated predictive capability back into the research agent to drive the next round of hypothesis generation and experimental-strategy optimization—with researchers confirming and reviewing at key checkpoints—forming the P0 self-evolving loop of "intelligent hypothesis → autonomous experiment → data feedback → operator evolution → intelligent re-exploration."

• P1 Domain Loop (cross-disciplinary deep exploration, multiscale transfer): Once initial leads and base models have been validated, RSI's cognitive agent adopts an intelligent exploration strategy coordinating "Exploration–Exploitation," distilling historical trial-and-error and experimental trajectories into a lightweight "offline replay simulator" for low-cost reasoning and pruning in a virtual latent space. It monitors gradient discrepancies across scales from micro to macro, dynamically schedules inverse-problem solving, and triggers online digital experiments for high-confidence candidate solutions, extending validation to heat-resistant monomer ratios, rubber particle-size distributions, processing-additive combinations, and comprehensive tests such as low-temperature impact and long-term aging—enabling data and experience to flow seamlessly among formulation screening, experimental validation and process optimization.

• P2 Industrial Loop (from laboratory to production line, a digital-twin closed loop): Even after an excellent formulation is validated in the laboratory, performance deviations often emerge when it is scaled up in reactors and twin-screw extruders or moved into plant production, owing to factors such as shear forces and raw-material batch variability. RSI extends to the production environment the cross-scale models it has accumulated in the domain, delivering predictions within seconds, while temperature, pressure, torque, shear-history and product-quality data generated by DCS/MES systems during production are continuously fed back into the R&D knowledge graph. Through industrial-data-driven operator retraining and model-deviation diagnosis, the platform achieves dynamic digital-twin calibration. By connecting real operating-condition feedback from the production side and channeling pilot- and production-scale data back to the R&D side for model correction, the P2 industrial loop bridges the "valley of death" between the laboratory and the plant floor.

Addressing LLM Hallucination: Deep Fusion of Physical AI and the Fabarta Context Graph

Unlike conventional tools that rely on the "probabilistic guessing" of general-purpose large models, the RSI platform is built on Fabarta's core Context Graph (enterprise knowledge graph) and physics-informed neural operator (PINO) models, ensuring rigor and traceability in scientific decision-making:

• Precise traceability and knowledge consolidation: RSI organizes raw monomers, formulation parameters, test results, phase-morphology images, patent literature and researchers' reasoning hypotheses into a unified, structured knowledge network. Researchers can trace across projects (for example, "retrieve the historical rubber particle-size characteristics of formulations that introduced N-phenylmaleimide while maintaining toughness"), and every conclusion is backed by explicit source data—eliminating the "model hallucination" of general-purpose AI in rigorous scientific domains.

• Hybrid physics-mechanism and AI modeling: The platform integrates equation-of-state, reaction-kinetics and simulation capabilities, embedding physical laws—such as conservation laws and constitutive relations—into operator training through a combination of soft and hard constraints, so that models remain physically consistent in data-sparse regions and deliver credible extrapolation. Computation becomes conversational, reasoning steps are visualized, and intermediate results can be re-examined.

• Bidirectional reinforcement between knowledge and models: The knowledge graph provides PINO with structured priors and training corpora, while PINO's predictions and discoveries are written back to the graph as "knowledge with confidence levels" and, once confirmed through experiments, consolidated into deterministic knowledge. The graph and the models feed each other and co-evolve, forming an enterprise-exclusive, non-replicable R&D intelligence asset.

Advancing the Closed-Loop Deployment of AI4S Across Industry, Academia, Research and Application

To date, Fabarta's AI4S solutions have been successfully deployed at leading central state-owned enterprises, national-level chemical industrial parks and multiple listed new-materials companies. Through joint laboratories, integrated dry- and wet-lab experimentation, and pilot-scale bases operated by government or enterprises, the RSI platform is accelerating the commercialization of catalyst optimization, polymer design and fine-chemical process transformation.

"In real industrial scenarios, AI must answer two questions: explainability and the cost of deployment. Through 'Graph Fine-tuning' and 'standardized components + scenario-specific configuration,' the RSI platform not only lowers the trial-and-error cost for enterprises adopting AI, but also elevates AI from mere literature Q&A to an AI R&D and production team covering molecular design, digital pilot-scale trials and energy-saving, cost-reducing production." — Andy Gao, Founder and CEO of Fabarta

Looking ahead, Fabarta will continue to center on research and production, further deepening its work in multi-agent collaboration, physics-informed AI hybrid modeling and its matrix of vertical-domain large models, and will remain committed to providing highly reliable, evolving enterprise-grade AI agent infrastructure for chemical-energy, new-materials, advanced-manufacturing and biopharmaceutical companies worldwide.

About Fabarta

Fabarta Co., Ltd. is an industry-leading provider of enterprise-grade AI agents and knowledge engines. Guided by its dual-engine strategy of "knowledge engine + large model," the company leverages its proprietary Graph Fine-tuning and physics-informed AI hybrid modeling technologies to solve the hallucination and inexplicability problems of general-purpose large models in industrial domains. Fabarta is deeply engaged in trillion-yuan-scale industries including fine chemicals, new materials and advanced manufacturing, and has built a complete AI4S deployment loop spanning "laboratory bench-scale research, digital pilot-scale trials and industrial mass production." The company is committed to helping manufacturing customers build smart factories based on four-dimensional ontology and RSI technology, opening the closed loop from research to production. As a strategic partner of leading value-chain anchor enterprises, Fabarta has achieved standardized, at-scale replication of more than one hundred mature agents at national-level chemical parks and among industry leaders, dedicated to embedding AI into the core value chain of the real economy and creating measurable industrial value.