Designing concrete that can survive an earthquake has long been a slow, fragmented craft, stitched together from laboratory trial mixes, expert judgment, and laborious finite element simulations. A new study published in AI in Civil Engineering reports that much of this process can now be handed over to a team of cooperating artificial intelligence agents. The system, built by researcher Artem Zaitsev, autonomously designs fiber-reinforced concrete formulations tailored to seismic conditions, and it does so while keeping the underlying physics firmly in the loop.
The multi-agent system, or MAS, targets the components that form the seismic load-bearing envelope of multi-storey buildings: shear walls, beams, and columns. Rather than acting as a general-purpose structural tool, the framework generates complete fiber-reinforced concrete specifications, choosing fiber type, geometry, and reinforcement ratios according to seismic inputs and project constraints that it parses automatically. These inputs include building typology, number of storeys, seismic zone classification, soil conditions, and local construction practices. The system then identifies where within each structural member additional fiber reinforcement is needed to improve energy dissipation, ductility, and post-cracking performance, and recommends reductions in conventional steel reinforcement where fibers can do the job.
The architecture combines two strands of artificial intelligence. Agentic AI provides distributed reasoning, coordination, and task orchestration across five specialized agents, while Physical AI embeds physics-informed models directly into the decision-making loop. The five agents divide the work: an Input Parser Agent extracts seismic and project parameters while checking compliance with codes such as ACI 318, ASCE standards, and Eurocode; a Material Designer Agent runs entropy-based stochastic micromechanics to predict concrete performance; a Microstructure Vision Agent analyzes volumetric micro-CT scans; a Manufacturing Coordinator Agent translates designs into fabrication-ready specifications; and a Validation Agent cross-checks designs against structural safety databases before approving export to external finite element software.
The reasoning backbone is built on two large language models, Qwen-30B-A3B and Phi-4-14B, grounded by agentic retrieval-augmented generation, a dynamic variant of retrieval-augmented generation in which agents autonomously trigger and refine knowledge retrieval based on their intermediate reasoning states. Agents communicate through an adapted version of the Agent2Agent protocol, an open standard using JSON RPC 2.0 and gRPC with TLS encryption, extended here with structural ontologies, uncertainty exchange, and dynamic task delegation. Every agent runs locally in Docker containers within Ollama on a workstation equipped with an RTX 4080 GPU, preserving data privacy while the LangGraph orchestration framework manages stateful, cyclic workflows with feedback loops.
The physics core rests on a modified maximum entropy principle. Instead of brute-force Monte Carlo sampling, the Material Designer Agent maximizes the Shannon entropy of the probability distribution over microstructural states, subject to physically measurable constraints such as fiber volume fraction, interfacial transition zone thickness, void fraction, and crack density. A key modification adds pairwise interaction descriptors, allowing the distribution to capture correlated effects like fiber clustering, fiber-aggregate overlap, and void adjacency to inclusions. The Lagrange multipliers are solved by Newton-Raphson iteration on a strictly convex dual objective, guaranteeing rapid convergence and enabling real-time performance. The resulting distributions feed homogenized stiffness tensors and cyclic energy dissipation predictions, which are carried into a reduced-order multi-degree-of-freedom model of the building shell.
Perception is handled by a 3D U-Net convolutional neural network trained on 196 volumetric micro-CT scans of basalt, steel, and carbon fiber-reinforced concrete specimens, each 512 by 512 by 1,700 voxels. The network segments the concrete into matrix, aggregates, fibers, interfacial transition zones, and voids, with a hybrid loss combining cross-entropy, Dice similarity, and physics-informed penalties that discard physically implausible geometries such as disconnected fibers. Fiber voxels make up less than two percent of the volume, a severe class imbalance addressed with weighted Dice loss and patch-based sampling. The trained model achieved Dice coefficients of 0.94 for aggregates and 0.87 to 0.89 for fibers, with segmentation performance varying by no more than plus or minus 0.02 in Dice across the three material types. Voids are modeled as triaxial ellipsoids rather than spheres, with their shapes extracted from segmented scans by principal component analysis.
The system was evaluated on a dataset of 3,000 simulated seismic design and material optimization scenarios accumulated over three years, spanning peak ground accelerations from 0.15 g to 0.45 g, multiple soil classes, frame shear-wall structural typologies, and stochastic variation in fiber type, volume fraction, aspect ratio, and rubber content. Benchmarking showed an average semantic retrieval latency of 530 milliseconds per agent query, a top-30 context relevance hit rate of 92.3 percent, and reasoning accuracy of 89.2 percent when cross-validated against verified seismic design workflows. When precomputed microstructural descriptors are available, the system bypasses segmentation and cuts a full design cycle from roughly 24 minutes to under 155 seconds; novel compositions requiring de novo characterization take a median of 22 minutes for a 1,700-slice reconstruction on the local workstation.
Structural-level validation was performed in ABAQUS under low-cycle, earthquake-representative loading, using quasi-static cyclic displacement protocols at a strain rate of 10⁻² per second with stress demands of 18 to 28 megapascals in critical hinge regions. Across basalt, steel, and carbon fiber-reinforced concrete, the system’s predictions tracked the median of three independent energy dissipation frameworks, with mean bias below 4 percent for all material types and root-mean-square errors of 50 to 80 kilojoules against dissipation magnitudes of 1.5 to 3.0 megajoules, roughly 2 to 5 percent relative error. Failure state classification reached an overall F1 score of 0.98, with shear walls slightly lower at 0.976, reflecting the complexity of shear-dominated damage. Fourteen independent structural engineers audited the workflow across 120 seismic loading scenarios, yielding a peer validation adherence score of 91.4 percent against their own manual finite element simulations and code compliance assessments.
The predicted energy dissipation patterns differed meaningfully by material. Steel fiber-reinforced concrete showed the highest dissipation, reaching roughly 2,500 to 2,950 kilojoules in beams, columns, and shear walls, attributed to the mechanical anchorage of hooked fibers and their resistance to crack widening. Basalt fiber-rubberized concrete dissipated 700 to 2,100 kilojoules in columns, concentrated sharply at the base, while carbon fiber concrete showed intermediate levels with more distributed patterns. In fragility analysis, a deterministic baseline neglecting microstructural variability produced a narrow collapse probability curve reaching about 72 percent at 2 percent drift, whereas the entropy-based posterior yielded a broader and more realistic envelope with a mean collapse probability of 80 percent at the same drift, better capturing rare but critical configurations such as void clusters near the interfacial transition zones.
The practical payoff is speed. A complete design cycle, including micro-CT segmentation, stochastic micromechanical convergence, and manufacturing-ready specification generation, takes 30 to 40 minutes per structural element, dropping to 7 to 12 minutes when descriptors are pre-indexed, with predictive bias increasing by less than 0.5 percent. Equivalent finite-element-driven design iterations require 2 to 8 hours of expert-led modelling, corresponding to an empirically observed reduction of approximately 75 percent in human-intensive design cycles. The authors note that accuracy remains dependent on training data quality and model calibration, and that full-scale structural modelling of the building shell lies beyond the study’s scope, with outputs exported instead to external finite element environments for independent verification. Future work points toward physics-informed large language models and large concept models that could embed fracture mechanics and fiber pull-out physics directly into inference, extending the framework to unseen material systems.
Subject of Research: An autonomous multi-agent AI system that designs seismic-resilient fiber-reinforced concrete using physics-informed stochastic modelling and micro-CT image segmentation
Article Title: Automating the seismic-resilient design of fiber-reinforced concrete using a physics-informed multi-agent system
Article References: Zaitsev, A. (2026). Automating the seismic-resilient design of fiber-reinforced concrete using a physics-informed multi-agent system. AI in Civil Engineering, 5(1), Article 20. https://doi.org/10.1007/s43503-026-00102-z
Image Credits: AI Generated
DOI: 10.1007/s43503-026-00102-z
Keywords: fiber-reinforced concrete, seismic design, multi-agent system, agentic AI, maximum entropy principle, micro-CT segmentation, 3D U-Net, energy dissipation, large language models, structural engineering, fragility analysis, Agent2Agent protocol
News Source: Violet Maxwell. (October 29, 2026). AI Agent Team Automates Earthquake-Resistant Concrete Design. Scienmag.
