OpenSees Seismic Structural Simulation from Nonlinear Fibres to ML-Based Damage Assessment
New1. Overview
Seismic risk studies must simulate thousands of buildings under dozens to hundreds of ground motions, and each building needs a nonlinear time-history run. That combination pushes the problem to the edge of what conventional engineering software can absorb. OpenSees was built to address exactly this tension: it makes the nonlinear finite element core scriptable and batch-callable, turning a single case into a parameter sweep.
Its provenance deserves precision. The source repository OpenSees/OpenSees was created on 12 September 2018, with copyright held by The Regents of the University of California and a copyright notice spanning 1999 to 2020. The footer of the official site opensees.berkeley.edu notes support from the National Science Foundation. The repository was last pushed on 6 November 2026, the default branch is master, the primary language is C++, and the project currently carries about 856 stars, 842 forks and 68 open issues, a scale that stands out among open-source structural simulation projects.
Two facts belong at the top of this entry because both change how the software may be used.
First, in the strict sense this is not a standard open-source license. The COPYRIGHT file at the repository root grants the Regents permission, without fee and without a written license agreement, for use, reproduction, modification and distribution of the software and its documentation by educational, research and non-profit entities for noncommercial purposes only, and for use, reproduction and modification by other entities for internal purposes only. Incorporating the software into products for commercial distribution requires contacting the University of California Office of Technology Licensing. The GitHub license field accordingly reports NOASSERTION. This differs materially from Apache-2.0 or MIT, and any commercial engagement must clear a compliance review first.
Second, the interface is layered. The Tcl interpreter is the historical main line, and the official documentation, user manual and the large majority of examples are written in Tcl. OpenSeesPy wraps the same kernel as a Python library in the repository zhuminjie/OpenSeesPy, created on 14 September 2018, currently at roughly 290 stars, 97 forks and 42 open issues, with default branch openseespy and last push on 28 October 2026. This Python layer is the key to connecting OpenSees to a physical-AI workflow, because it turns parameter sweeps, hyperparameter search and bulk data generation into a few lines of script instead of a hand-written loop inside a Tcl interpreter.
2. How It Works
OpenSees is a command-stream analysis engine: the user defines a model, materials, loads, the solution procedure and the output in sequence, and the commands execute in order and then release memory. The modelling paradigm can be summarised as elements as carriers, materials as the core, and recorders as the exit.
The material layer carries most of the expressive power. Under SRC/material the repository splits into three groups. The uniaxial directory holds 344 source files and roughly 164 implementation classes, covering steel (Steel01, Steel02, SteelMP, Steel4), concrete (Concrete01, Concrete02, Concrete04, Concrete05, Concrete06, Concrete07, ConcreteCM and the TDConcrete family), soil liquefaction (PressureDependentMultiYield, PressureIndependentMultiYield, PM4Sand, PM4Silt, PyLiq1, QzLiq1, TzLiq1), damage and degradation (DuctileFracture, SteelFractureDI, SimpleFractureMaterial, DegradingPinchedBW), and isolation devices such as ElastomericBearing, FrictionSpringDamper and the KikuchiAikenHDR and KikuchiAikenLRB constitutive models paired with isolation bearings. The nD directory holds 165 source files and about 80 multiaxial classes including DruckerPrager, capped plasticity, J2Plasticity, ManzariDafaliasASDSoil, the CycLiqCP family for saturated soil cyclic plasticity, PlasticDamageConcrete3d and fibre-based plastic damage materials. The section directory holds 88 source files and about 41 section classes, spanning FiberSection2d and 3d, NDFiberSection, ElasticMembranePlateSection, LayeredShellFiberSection and SectionAggregator.
The element layer is equally broad. The 46 subdirectories under SRC/element include dispBeamColumn, forceBeamColumn, nonlinearBeamColumn, elasticBeamColumn, mixedBeamColumn, beam2d and beam3d, truss, brick, fourNodeQuad, tetrahedron, shell, masonry, joint, zeroLength, twoNodeLink, elastomericBearing, frictionBearing, catenaryCable, mvlem, absorbentBoundaries, plus fluid-structure interaction families and the PFEMElement and IGA extensions.
The solution layer splits into integrators, linear equation solvers, constraint handlers and iteration algorithms. The integrator directory contains the transient schemes Newmark, CentralDifference, HHT, GeneralizedAlpha, TRBDF2, TRBDF3, ExplicitDifference, ExplicitBathe, AlphaOS, Houbolt, KRAlphaExplicit, Collocation and ParkLMS3, together with the static and path-dependent forms LoadControl, DisplacementControl, StagedLoadControl, ArcLength, EQPath and HarmonicSteadyState. The linear solvers provide BandGeneral, BandSPD, FullGeneral, ProfileSPD, SparseSYM, SuperLU, Umfpack, Mumps, plus AMD and RCM preconditioning. The v3.8.0 release notes record a new eigenvalue solver based on the LAPACK routine dsygvx and a Sparse interface for Python-side linear and eigenvalue solvers.
Seismic input is handled by the groundMotion, plainGroundMotion and interpolatedGroundMotion commands combined with time series types for constant, path, triangular, pulse and trigonometric patterns. Output is provided by NodeRecorder, ElementRecorder, EnvelopeNodeRecorder, GmshRecorder, PVDRecorder and MPCORecorder, with v3.8.0 adding a VTK_HDF recorder.
3. Scope and Assumptions
The usable scope of OpenSees can be read back from its modelling orientation, which is an advantage in most settings and a limitation in others.
First, it centres on concentrated plasticity or fibre discretisation. Component-level nonlinearity is expressed mainly through section fibre discretisation or plastic hinges. This system is sufficient in accuracy for flexure-dominated components, but strong-axis shear, joint-region failure, rebar buckling and bond-slip require additional specialised elements or zero-length elements, and the workload rises sharply.
Second, it is not a general multi-physics model. The fluid domain, soil-structure interaction, interface friction and foundation connections each have dedicated implementations, but they exist to serve structural response and do not attempt general multi-physics coupling. Resolving bidirectional fluid-structure coupling, thermo-mechanical coupling or mesoscopic concrete cracking requires building additional models inside the framework.
Third, the calibration cost of material models cannot be skipped. Every constitutive model in the library has its intended application and its calibration requirements, and the three degradation parameter families, pinching, stiffness degradation and strength degradation, can shift the response of the same component by a factor of several. Applying literature defaults directly yields a formally correct time history that cannot be trusted.
Fourth, the licensing boundary is a hard constraint, as described earlier. Educational and research use is free, while commercial productisation requires going through the licensing process, and this single fact determines the tool's position inside a commercial software stack.
Fifth, reliability depends on sensitivity to numerical parameters. Convergence tolerances, integration step size and the constraint handling choice of Plain, Penalty, Lagrange or Transformation all affect results, and in the strongly nonlinear stage close to failure, improper settings cause the time history to terminate early or diverge.
4. Getting Started
The documentation repository OpenSees/OpenSeesDocumentation, at about 67 stars, lays out the full path. The user manual is organised in five chapters: analysis, material, model, output and section. The analysis chapter covers analysis types, integrators, constraints, algorithms, numberers and linear systems across 42 entries. The model chapter holds 49 entries covering nodes, elements, load patterns, time series, ground motions, damping and geometric transformations. The material chapter lists 42 uniaxial material pages, 23 multiaxial material pages and the section descriptions. The output chapter lists seven recorders.
For Python users, the openseespy package on PyPI is the shortest route. The current version is 3.8.0.0, uploaded on 18 March 2026, requiring Python 3.10 or later and distributed through three platform packages named openseespylinux, openseespywin and openseespymac. PyPI hosts 65 release versions in total, with the history traceable back to the 3.5.1.x and 3.6.0.x series, broadly tracking the 3.6, 3.7 and 3.8 kernel cadence.
A recommended practical sequence has five steps. First, run the smallest closed loop using an official example: the EXAMPLES directory in the repository and the user/examples section of the documentation site provide many Tcl cases, so pick the simplest single-degree-of-freedom or single-bay frame case and confirm the environment runs it. Second, rewrite the same case in Python, mapping Tcl commands one to one onto OpenSeesPy functions to establish a semantic correspondence. Third, complete the output by exporting the engineering demand parameters of interest, such as interstorey drift ratio, roof acceleration, residual deformation and member plastic rotation, as tabular records. Fourth, introduce real ground motions, reading acceleration records through uniformExcitation combined with a time series, keeping units and damping assumptions consistent. Only the fifth step is the bulk sweep, wrapping the third step's output interface into a function and running the combinations of parameters and motions in a loop or through parallel tooling.
The community rules are worth noting. The repository README states plainly that the repository does not accept direct writes, that pull requests are the only accepted contribution channel, and that contributors should fork to their own account and follow a forking workflow. Modelling questions raised as issues are closed and redirected to the OpenSees message board or the Facebook group.
5. Comparison with Alternatives
Placed in the landscape of structural simulation, OpenSees is positioned by several reference points.
Against commercial finite element packages such as ABAQUS, ANSYS, ETABS and SAP2000, OpenSees offers scripting and batch capability plus a material and element library tailored to earthquake engineering. It lacks a complete pre- and post-processor, general multi-physics coupling and commercial-grade quality assurance, and has no systematic third-party validation scheme. What is distributed on the official site is a Windows binary plus macOS builds for both Intel and ARM, with a Tk-toolkit version retained for interactive use, which leaves the user experience some distance from commercial software.
Against peer open-source structural tools, OpenSees leads in community size and literature depth: the counts listed in the official manual and documentation pages, 42 uniaxial material pages, 23 multiaxial material pages and 39 element pages, reflect coverage accumulated over many years.
Against numerical tools in the wider Python ecosystem, OpenSees occupies the niche of a domain-specific high-fidelity solver. It offers no automatic differentiation and no native integration with deep learning frameworks, and the Sparse interface for Python-side solvers only arrived in v3.8.0. The typical division of labour is that OpenSees generates the labels while PyTorch, PhysicsNeMo or a purpose-built network learns from them, with the interface being data files and call scripts.
Against machine-learning damage assessment approaches based on imagery or sensor streams, the OpenSees route is simulation-driven. Its advantage is coverage of buildings with no measured data and access to the full response history rather than only a damage label. Its cost is that every sample costs far more compute than those alternatives, which makes surrogate models and order reduction more necessary rather than less.
6. Benchmarks and Evidence
The credibility of OpenSees rests on a chain of publicly traceable literature.
The foundational paper for the framework is McKenna's OpenSees: A Framework for Earthquake Engineering Simulation, published in Computing in Science and Engineering in 2011, volume 13 issue 4, pages 58 to 66. The formal reference for the Python layer is Zhu, McKenna and Scott's OpenSeesPy: Python library for the OpenSees finite element framework, published in SoftwareX in 2018, volume 7, pages 6 to 11.
A set of citable results addresses individual capabilities. In October 2025, An OpenSEES graphical user interface for structural dynamics and earthquake engineering instruction appeared in Scientific Reports volume 15, pointing to tooling for teaching. In 2026, An Integrated Open-Source Workflow for OpenSees was published in Procedia Structural Integrity volume 78, pages 1497 to 1504, discussing how OpenSees fits into an overall open-source workflow. The same year, opsvis: pre-, post-processing and visualization Python module for OpenSeesPy appeared in SoftwareX volume 34, article 102615, closing the visualisation gap on the Python side.
The version cadence is itself evidence. The repository has 11 releases: v3.0.0 on 5 November 2018, V3.0.2 on 29 January 2019, v3.1.0 on 22 November 2019, v3.2.0 on 25 March 2020, v3.3.0 on 17 May 2021, v3.4.0 on 9 December 2021, v3.5.0 on 11 May 2023, v3.6.0 on 31 January 2024, v3.7.0 on 30 July 2024, v3.7.1 on 20 January 2025, and v3.8.0 on 18 February 2026. The v3.8.0 notes list 16 new features and improvements, concentrated on deriving the acceleration series from the displacement series in GroundMotion, the VTK_HDF recorder, equation constraints, the LinearElasticGGmax and Equivalent Linear materials, the ACI209 creep and shrinkage uniaxial material, revised geometric transformations, surface loading for shell elements, new explicit integrators, and substantial updates to ASDPlasticMaterial3D and HystereticSM.
On ecosystem scale, verifiable quantitative indicators include six public repositories in the organisation: besides the main repository, the documentation repository at 67 stars, the examples repository at 6 stars, and the MUMPS solver and METIS graph partitioning submodules brought in under an MIT license. The material and element directories accumulate source files in the hundreds.
7. Role in Physics-AI Workflows
For physical-AI practitioners, OpenSees occupies the same position as the other solvers covered here, a scriptable ground-truth generator, with one distinguishing feature: its output is a complete time history rather than a scalar or a field distribution.
Generating surrogate training data is the most direct use. The inputs are discrete structural parameters covering storey count, span, member sections, reinforcement ratios, constitutive degradation parameters and damping ratio, together with ground motion records. The outputs are member force and member deformation time series. The typical training target maps ground motion plus structural parameters onto peak interstorey drift or a damage index, a mapping for which measured buildings rarely provide sufficient coverage, while simulation can generate any combination on demand.
Real-time damage assessment is the second use. The 2019 arXiv paper on intensity-based feature selection for near real-time damage diagnosis of building structures is an instance of this direction, using ground motion intensity metrics as features to achieve quasi-real-time diagnosis. Regional building portfolio assessment depends on the same capability: generate a large batch of cases with OpenSees, train a surrogate, then let the surrogate complete the assessment of tens of thousands of buildings in seconds.
Parameter inversion and uncertainty quantification is the third use, and it is where OpenSees differs from other solvers. Because it outputs complete force-deformation relations, it can plug directly into Bayesian calibration frameworks and back out constitutive parameters and boundary conditions from observations through a likelihood function. The 2023 Springer LNCE paper on sensitivity analysis and Bayesian calibration of OpenSees models using quoFEM links OpenSees to a sensitivity and Bayesian calibration toolchain, and a 2026 Infrastructures article integrates Bayesian inference into structural parameter estimation with a Python approach.
A more forward-looking direction connects large language models directly to the modelling workflow. The April 2026 arXiv paper on automating structural analysis across multiple software platforms using large language models and the March 2026 paper on a multi-agent architecture to reduce hallucinations in multi-step structural analysis both treat the OpenSees command interface as an action space callable by a language model, where the deterministic syntax of a command stream fits naturally as an action definition for tool calling.
Placed in the wider physical-AI ecosystem, OpenSees is a structural-side solver alongside bridge structural analysis, complementary to the inversion methods of OpenFWI seismic inversion and forming an upstream-downstream chain with topology optimization. The surrogate modelling and reduced-order collection and the Bayesian inversion and uncertainty quantification collection provide methodological references, and the numerical methods and solvers collection supports comparison of how solvers are organised across physical domains.
8. Limitations and Failure Modes
Several situations should be avoided or routed to other tools.
Situations requiring commercial certification and auditable provenance are a poor fit. Engineering practice often demands a traceable computation and third-party validated software, and the licensing and validation ecosystem around OpenSees is limited in that respect.
Situations requiring full multi-physics coupling exceed the scope of the framework.
Situations requiring mesoscopic concrete cracking are poorly served, because concrete constitutive behaviour is macroscopic and crack width or local spalling needs dedicated models.
Situations requiring long-term creep and shrinkage assessment deserve caution. Although v3.8.0 added an ACI209 creep and shrinkage uniaxial material, this remains a supplementary capability in the modelling chain, and long-term deformation of concrete structures is usually handled by dedicated tools.
Situations demanding a mature GUI and post-processing experience are also a mismatch. The official paths are Tcl and Python scripting, with visualisation depending on third-party modules such as opsvis or on your own handling.
For physical-AI workers in particular, two failure modes deserve emphasis. The first is training on uncalibrated constitutive parameters. The OpenSees material library is large, and the difference in time history produced by different degradation models for the same member can exceed the difference between two different structures, so training data that is not uniform at the calibration level teaches the network the conventions of the numerical setup rather than physical reality. The second is treating numerical artefacts as physical response. The combination of convergence tolerance, integration step size and constraint handling affects peak response, and when the target quantity is peak interstorey drift, a quantity sensitive to step size, step-size convergence verification is mandatory, otherwise the trained model carries a systematic bias when applied to real structures.
One engineering constraint also applies: OpenSees memory grows quickly with model size, so large-scale portfolio simulation requires a trade-off between model simplification and fidelity, and the simplification scheme must match the intended use.
9. Version and Source Notes
The factual basis of this entry is recorded below for later alignment.
The repository is at https://github.com/OpenSees/OpenSees with default branch master, primary language C++, and last push on 6 November 2026. The documentation site is https://opensees.github.io/OpenSeesDocumentation with sources in the OpenSees/OpenSeesDocumentation repository. The official site is https://opensees.berkeley.edu/ , whose download page states the current version is 3.8.0 and offers a Windows 64-bit binary, x86_64 and ARM macOS builds, and a Tk version. The Python distribution is published on PyPI as openseespy with sources at https://github.com/zhuminjie/OpenSeesPy .
The licensing basis follows the COPYRIGHT file at the repository root: copyright held by The Regents of the University of California from 1999 to 2020, free for noncommercial use by educational, research and non-profit entities, internal use only for other entities, and commercial redistribution requiring an application to the University of California Office of Technology Licensing. The GitHub API reports the license as NOASSERTION, consistent with that file.
The version timeline follows the GitHub Releases list, which holds 11 releases with the latest being v3.8.0 on 18 February 2026. The Python package uses a four-segment version number, currently 3.8.0.0, which tracks the kernel version without being perfectly synchronised, so the two must be distinguished when citing.
Six public repositories sit under the organisation: OpenSees at 856 stars, OpenSeesDocumentation at 67, OpenSeesExamples at 6, SimCenterBootcamp2020, MUMPS under an MIT license, and METIS. The contribution flow for the main repository is to fork and submit a pull request, with direct writes refused, and modelling questions are handled outside the issue tracker.
One boundary on the counts deserves stating: the material and element figures come from tallying source files in the SRC/material and SRC/element directories, which is an approximation of code scale and does not match the number of entries listed with dedicated pages in the official manual. When citing a specific capability, the manual pages are the reference.
10. References
- OpenSees source repository, https://github.com/OpenSees/OpenSees , accessed 12 November 2026.
- OpenSees official documentation site, https://opensees.github.io/OpenSeesDocumentation , with the user manual and developer guide.
- OpenSees official site and download page, https://opensees.berkeley.edu/ , current release version 3.8.0.
- Repository COPYRIGHT file, https://github.com/OpenSees/OpenSees/blob/master/COPYRIGHT , stating copyright ownership and permitted uses.
- McKenna, OpenSees: A Framework for Earthquake Engineering Simulation, Computing in Science and Engineering, volume 13 issue 4, pages 58 to 66, 2011, DOI 10.1109/mcse.2011.66.
- Zhu, McKenna and Scott, OpenSeesPy: Python library for the OpenSees finite element framework, SoftwareX, volume 7, pages 6 to 11, 2018, DOI 10.1016/j.softx.2017.10.009.
- Kokot, opsvis: pre-, post-processing and visualization Python module for OpenSeesPy, SoftwareX, volume 34, article 102615, 2026, DOI 10.1016/j.softx.2026.102615.
- Ercolessi et al., An Integrated Open-Source Workflow for OpenSees, Procedia Structural Integrity, volume 78, pages 1497 to 1504, 2026, DOI 10.1016/j.prostr.2025.12.191.
- Yazici and Larouzi, An OpenSEES graphical user interface for structural dynamics and earthquake engineering instruction, Scientific Reports volume 15, 2025, DOI 10.1038/s41598-025-17632-8.
- Sensitivity Analysis and Bayesian Calibration of OpenSees Models Using quoFEM, Lecture Notes in Civil Engineering, Springer, 2023, DOI 10.1007/978-3-031-30125-4_6.
- Integrating Bayesian Inference into Structural Parameter Estimation: A Python-Based Approach, Infrastructures, volume 11 issue 8 article 291, 2026, DOI 10.3390/infrastructures11080291.
- OpenSeesPy project page on PyPI, https://pypi.org/project/openseespy/ , current version 3.8.0.0.
- OpenSees GitHub releases page, https://github.com/OpenSees/OpenSees/releases , carrying release notes from v3.0.0 through v3.8.0.
- arXiv 2604.09866, Automating Structural Analysis Across Multiple Software Platforms Using Large Language Models, 2026.
- arXiv 2603.07728, A Novel Multi-Agent Architecture to Reduce Hallucinations of Large Language Models in Multi-Step Structural Modeling, 2026.
- arXiv 2605.09175, An Open-Source Framework for Coupled Vehicle-Bridge Interaction Analysis Using OpenSees, 2026.
- arXiv 1910.11240, Intensity-Based Feature Selection for Near Real-Time Damage Diagnosis of Building Structures, 2019.
- arXiv 2312.09450, Performance-Based Optimization of 2D Reinforced Concrete Moment Frames through Pushover Analysis and ABC, 2023.
- Modeling viscous damping in nonlinear response history analysis of buildings for earthquakes, Earthquake Engineering and Structural Dynamics, 2015, DOI 10.1002/eqe.2622.
- A Practical Guide to OpenSees, Frontier Research in Computation and Mechanics, World Scientific, 2023, DOI 10.1142/11522.