Multi-Agent Based Automation of Geotechnical Risk Assessment

Manual geotechnical analysis is a sequential bottleneck, often causing costly redesigns after structural engineering begins. A custom workflow automates the ingestion of borehole logs, CPT data, and geophysical surveys into a unified soil model. Agents run preliminary finite element analysis against proposed foundation loads, flagging settlement or bearing capacity issues before detailed design locks in. This shifts risk identification from the construction documents phase to early site investigation, preventing weeks of rework.

Compiling geotechnical baseline reports (GBRs) involves manually correlating data from multiple consultants and formats. A custom agentic workflow automates this synthesis: one agent extracts stratigraphy from logs, another calculates soil parameters, and a third drafts risk narratives and visualizations. The system generates a draft GBR with embedded risk heatmaps, ready for senior review. This accelerates the critical path to foundation design and procurement.

Unquantified ground risk forces contractors and owners to pad bids with large contingency allowances. A custom workflow quantifies risk probabilistically, modeling scenarios for excavation stability, dewatering needs, and pile installation challenges. By providing a data-driven risk score and mitigation recommendations, the workflow allows for more precise contingency planning. This translates directly into more competitive, confident bids and protects project margin from unforeseen ground conditions.

The greatest cost of geotechnical uncertainty manifests as field changes—unexpected rock, water ingress, or slope failure. A custom workflow's predictive modeling identifies high-probability risk zones (e.g., liquefaction potential, weak soil lenses) and layers them directly onto the site BIM. This allows the construction team to pre-plan mitigation strategies, such as alternative excavation support or ground improvement, during the preconstruction phase. The result is fewer surprises, fewer change orders, and smoother site operations.

Traditional geotechnical data is locked in static PDF reports. A custom workflow builds a living geotechnical digital twin—a queryable, 3D model of subsurface conditions linked to all raw data and analysis assumptions. This becomes a single source of truth for the entire project lifecycle, from design and litigation support to future asset management. It provides full auditability for decisions, supports insurance and dispute resolution, and adds long-term asset value.

A phased implementation starts with a pilot on a single site or project phase. Architecture involves: 1) Data Ingestion Agents to process .LAS, .DXF, and PDF survey data; 2) Modeling Agents to run Plaxis or RS2 simulations via API; 3) Orchestration via LangGraph to manage the analysis sequence; and 4) Integration with Autodesk Civil 3D or Bentley gINT for BIM output. A pilot typically demonstrates quantified risk reduction and process acceleration within one quarter, proving the ROI for a full rollout.

This agent orchestrates the ingestion and normalization of fragmented geotechnical data—borehole logs, CPT results, seismic surveys, and lidar topography—into a unified geospatial model aligned with the site BIM. It handles disparate formats (PDFs, CAD, .LAS) and resolves coordinate system conflicts, creating a single source of truth for subsurface conditions. Implementation uses LangChain for document parsing and GIS libraries (e.g., GDAL) for spatial alignment, feeding a graph database that links soil properties to BIM elements.

A specialized agent triggers and manages cloud-based FEA simulations (e.g., using Plaxis, RS2, or custom solvers) based on the fused geospatial model. It parameterizes models for settlement analysis, slope stability, and excavation support design, running hundreds of scenarios to quantify probabilistic risk. The agent monitors simulation completion, extracts key outputs (factors of safety, displacement contours), and layers results back into the BIM as risk heatmaps. This replaces manual, batch-run modeling with an event-driven, scalable analysis pipeline.

This agent interprets FEA outputs and historical incident data to generate actionable risk assessments. It classifies areas as high, medium, or low risk for excavation collapse, bearing capacity failure, or liquefaction. It then drafts preliminary mitigation recommendations—such as suggesting deep foundations over spread footings or recommending soil improvement techniques—and attaches these notes to relevant BIM components. The logic combines rule-based thresholds with LLM summarization for clear, context-aware reporting to engineers.

The workflow's output layer automatically pushes risk models and recommendations into the live BIM environment (e.g., Autodesk Revit, Bentley OpenBuildings) via API. Geotechnical risk surfaces are visualized as color-coded volumes within the model, and recommended foundation types or excavation support systems are suggested as parametric families. This creates a direct, visual link between subsurface uncertainty and design decisions, enabling real-time trade-off analysis during coordination meetings. Integration requires custom Dynamo or Forge scripts to manage bidirectional data flow.

A critical control point where all high-risk findings and automated recommendations are routed to a geotechnical engineer for review before they influence design. The agent packages the underlying data, simulation assumptions, and confidence scores into a review dashboard within the CDE (e.g., BIM 360). The engineer can approve, modify, or reject recommendations, with all decisions logged for auditability. This gate ensures professional liability is managed while still leveraging automation for the heavy lifting of data processing and scenario generation.

Deployment typically follows a pilot-on-one-site pattern. The core orchestration is built with LangGraph, managing agent handoffs and state. Agents are containerized for scalability, connecting to the CDE (like BIM 360) as the system of record. Legacy data ingestion often requires browser automation (Playwright) for portals without APIs. The entire stack is monitored for data quality drift (e.g., anomalous borehole data) and simulation runtime exceptions, with alerts routed to a devops team. Phased rollout focuses on one risk type (e.g., settlement) before expanding to full suite.