Computational Modeling of Infrastructure Materials in the Era of Artificial Intelligence 

Gianluca Cusatis, F.EMI, M.ASCE 
Professor of Civil and Environmental Engineering and Mechanical Engineering
Northwestern University


Abstract

This presentation offers a reflection on the status of computational modeling of concrete in a new era of advanced computational technologies readily available to researchers. It will first review the successes and the drawbacks of the last 25 years of high-fidelity computational modeling for concrete at various length scales and it will propose a vision for the future developments of the field. The presentation will discuss various computational approaches, but it will primarily highlight the benefits and superior modeling capabilities of discrete models, such as the Lattice Discrete Particle Model (LDPM). LDPM simulates accurately heterogeneous materials by directly linking the discreteness of the formulation with material heterogeneity at various length scales. In the last few decades, LDPM has been adopted successfully to simulate quasi-brittle failure and fracture under multiaxial loading conditions and dynamic loadings; to capture the multiphysical behavior of concrete and other quasibrittle materials under varying temperature and relative humidity; and to predict the response of reinforced concrete structural elements.

The second part of the presentation will discuss recent work coupling LDPM with Machine Learning (ML) techniques for several applications, including fast parameter identification, fragmentation characterization, and macroscopic constitutive modeling from fine-scale computational data. For parameter identification, the presentation will show results obtained by exploiting ML to obtain metamodels for typical laboratory tests for concrete, namely unconfined compression test, hydrostatic test, and three-point bending test. By using these ML metamodels the computational cost for a typical LDPM parameter optimization task is reduced by several orders of magnitude without sacrificing accuracy in the optimized parameters. The second ML study focuses on predicting and quantifying fragmentation phenomena under high impulsive dynamic loading via an innovative unsupervised learning clustering technique to identify and characterize mass and velocity of fragments. Results show excellent agreement with some available experimental data. A third ML application that will be discussed is relevant to the formulation of a macroscopic tensorial constitutive equation via the ML interpolation of a large database of homogenized LDPM response obtained with randomly generated strain histories featuring complex loading-unloading-reloading histories. The study highlights the limitations of methods that have been proposed in the literature but that have been only applied to relatively simple material behavior such as elasticity and classical plasticity. Finally, the presentation will conclude by discussing some current and, most importantly, potential future work adopting disruptive computational technologies that are expected to transform material modeling research and accelerate new discoveries. We will show the results of an exercise in which we challenged a certain class of ML methods to discover the non-locality of strain-softening response of concrete specimens subject to tensile fracture. This simple exercise demonstrates the superiority and accuracy of ML methods to analyze and interpolate data as well as to discover hidden information in data. At the same, it also highlights the inherent difficulty of providing physical interpretation of ML discoveries and the danger of using ML methods to extrapolate outside the available data.

Biography

Professor Gianluca Cusatis is a faculty member of the Civil and Environmental Engineering Department at Northwestern University that he joined in August 2011. Prior to joining Northwestern, Professor Cusatis worked at Rensselaer Polytechnic Institute for 6 years. He obtained his "Laurea" degree and his PhD degree in structural engineering from Politecnico Di Milano (Italy). His teaching experience covers several courses in structural mechanics and mechanics of materials typically offered in civil engineering curricula. His research expertise is relevant to experimental, computational and applied mechanics, with emphasis on heterogeneous and quasi-brittle infrastructure materials. His work on constitutive modeling of concrete through the adoption of the so-called Lattice Discrete Particle Model (LDPM), one of the most accurate and reliable approaches to simulate failure of materials experiencing strain-softening, is known worldwide. In addition, recent work on waterless concrete for Martian constructions has received widespread attention in the technical community and in the media. Under the sponsorship of several agencies, his current research focuses on formulating and validating multiscale and multiphysics computational frameworks for the simulation of large-scale problems dealing with a variety of different applications including, but not limited to, infrastructure aging and deterioration, structural resiliency, and response of materials and structures to natural and man-made hazards.

Professor Cusatis is a member of the American Society of Civil Engineers (ASCE) and the American Concrete Institute (ACI) and is active in several technical committees. He held leadership positions at Northwestern University (MMS Program coordinator and Director of Graduate Studies), ASCE EMI (Board of Directors), ACI 446 (Chair), ACI 209 (Chair), IA-ConCreep (President), and IA-FraMCoS (Treasurer). In 2018, he was awarded the EMI Fellow membership grade and in 2025 was elected Fellow for ACI. In 2020 he founded Cusatis Computational Services (CCS) Inc to pursue translation of fundamental research into practice.

From 2021 to 2024, he was on leave from teaching and service responsibilities at Northwestern University to serve as Program Director for the CMMI ECI program at the National Science Foundation. Professor Cusatis also served as CMMI ECI Expert from March 2024 to February 2025.