Artificial intelligence
The group develops AI tooling for its engineering work. Deterministic solvers perform the calculations and the language model handles routing, decomposition, and explanation. It does not compute values that a structure depends on.
Principles
Calculations come from the solver
Every number in a deliverable comes from running the solver in the current session rather than from memory or a stored file. The language model routes, decomposes, and translates, and does not compute safety-critical values.
No guessing
A standard, code, or test method is named only when that name appears in the source, with its defining parameters cross-checked first. Where a value is absent from the source, the gap is stated rather than filled in.
Verification against published results
Each computation module carries tests against published results, and outputs are attributable to their inputs, solver version, and run.
Human review before adoption
Model output is not adopted into the group's working procedures without human approval, and document retrieval is not treated as verified knowledge. Digitizers, crack counters, and fitting routines stop at a review step before a number is accepted.
Why this matters in engineering
A structural calculation is checked by a licensed engineer who carries the liability for it. That requires the same input to give the same answer every time, and every number to be traceable to the script, version, and inputs that produced it. A language model that estimates a residual strength cannot be checked in that way, so these systems do not allow it to.
Tools
Mechanics engines
in use
Section analysis engines
Closed-form strain-compatibility solvers for the flexural response of fiber reinforced and hybrid reinforced concrete. Multi-linear tension and compression laws with any number of segments, the quadratic neutral-axis solver for a fiber matrix carrying discrete rebar, layered and sandwich sections, FRP or TRC strengthened sections, and crack localization for inelastic deflection.
in use
Inverse analysis pipeline
Takes measured flexural test curves and recovers the tension law behind them. Batch auto-fit of four-point and three-point bending data, notched-beam crack-width inverse analysis on the EN 14651 CMOD basis, and a campaign runner that drives the engines over a list of validation cases, stores every result with its provenance, and re-renders the paper figures from the stored output.
in use
Fiber pullout and fracture mechanics
Single-fiber pullout with shear-lag interface mechanics and R-curve debonding, from the classical two-zone model to a closed-form generalization for an n-segment interface law, with every equation cited to its source publication and equation number. Alongside it, crack band theory, the size effect law, and first-crack snap-through instability in ASTM C1609 testing.
Design and code
in use
Design code assistants
Reference assistants over the standards the group works in, answering from the governing clause rather than from recall, across ACI 544 guidance for fiber reinforced concrete, elevated and ground-supported fiber slabs, precast tunnel segments, fib Model Code 2010 residual strength classes, and ACI 318 based design. The assistants are internal tools; the standards themselves remain the property of their publishers.
in use
Three-code slab calculator
Runs a steel fiber reinforced concrete slab through three independent design bases from one input set and reports them side by side, covering the North American, UK and European, and Swedish approaches. Punching shear, service stress, and crack width are checked in each basis, and safety formats are never mixed across codes inside a single check.
Simulation
in use
Finite element modeling
Member-level and continuum analysis of reinforced concrete, FRC, and UHPC, built and debugged against a searchable local index of the finite element documentation rather than from guesswork. The finite element layer never invents a constitutive law. It consumes the tension curve produced by the closed-form engines or measured in a test.
Data recovery
in use
Curve and figure digitization
Recovers the numbers behind published figures. Vector extraction from born-digital PDF figures with axis calibration, a separate path for scanned pages with tilted scans and dashed lines, and a third for off-axis photographs of a screen that corrects real perspective and lens distortion. Every path ends with an overlay of the trace on the original for human approval.
in use
Measurement and crack quantification
Digital image correlation workflows for crack width, crack spacing, strain field, and neutral axis migration on cementitious specimens under monotonic and fatigue loading. Radial and flexural crack counting on panel photographs and published crack maps, reported against reinforcement ratio, fiber dosage, or boundary condition, and stopping at a verification gate before any count is quoted.
Writing and figures
in use
Figures and technical reporting
A locked house-style plotting library for moment-curvature, load-deflection, stress-strain, and R-curve figures, a matching library for free-body diagrams, and calculation reports in which every equation is an editable native object authored in LaTeX rather than a flattened image, following the formula then substitution then result pattern.
in use
Literature search and synthesis
Full-text search and catalog maintenance over the group's research library, with duplicate detection and structured bibliographic fields, plus a multi-perspective research pipeline that runs question generation, retrieval-grounded drafting, a citation audit, and an explicit gap analysis.
Governance
in use
Verification and review agents
The enforcement layer. A senior review agent audits drafts the way a human collaborator would, checking audience fit as well as the formal rules. A pre-deploy checker audits any tool that emits engineering numbers against a discipline distilled from a real project post-mortem, covering adversarial fixtures, sanity bounds written as code, cross-checks against authoritative references, and refusal to guess.
research
Structural reliability solver
A First-Order Reliability Method solver verified against published benchmark cases, with SORM, Monte Carlo, reliability-based calibration of partial safety factors, and Bayesian updating. Its declared scope explicitly excludes domains it has not been verified for, so that it refuses rather than improvises outside its boundary.
Architecture research
research
Domain brain architecture
A written specification for persistent, narrow-scope engineering intelligence: a router that delegates but is forbidden from answering, versioned domain systems that each know one field and refuse questions outside a declared scope, a doctrine layer of reviewed truths, a deterministic solver layer that owns all safety-critical computation, benchmarks with tolerance windows, provenance carrying versions and run identifiers, and a human review layer that approves promotion.
SMC Labs
A research lab spun out of the group, building verified engineering intelligence systems for high-consequence technical decisions.
SMC Labs, for Structures, Materials and Composites, applies the same discipline outside the university. Its central idea is the artificial domain expert, a narrow system that performs one engineering check or code calculation end to end. It takes structured inputs, runs a deterministic solver against verified code logic, and returns a result with the calculation steps shown and the governing clause named. The language model translates that result into plain English and holds none of the engineering authority.
The constraint the team works under is that building one of these requires someone who understands the physics behind a code clause rather than only its words. Their phrase for it is that only a human domain expert can create an artificial domain expert.
The first deployed tool of this kind takes a dropped pin and a drawn parcel boundary and returns a code-referenced feasibility report with a foundation recommendation, cut and fill estimate, structural loads, and a long-range cost projection, assembled from public government GIS layers. A rule engine makes every engineering decision in it. It won its track at HackASU 2025.
Founded by Devansh Patel, Chidchanok Pleesudjai, Barzin Mobasher.