Application of data-driven neural networks to bio-inspired lattice design and prediction of multiphysics solution fields

Kushwaha, Shashank

Permalink

https://hdl.handle.net/2142/132763 Copy

Description

Title

Application of data-driven neural networks to bio-inspired lattice design and prediction of multiphysics solution fields

Author(s)

Kushwaha, Shashank

Issue Date

2025-11-24

Director of Research (if dissertation) or Advisor (if thesis)

Jasiuk , Iwona

Doctoral Committee Chair(s)

Jasiuk , Iwona

Committee Member(s)

• King , William P
• Koric, Seid
• Krishnan, Girish

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Deep operator networks
• Neural network surrogate modeling
• Additive manufacturing
• Multiphysics modeling
• Finite element analysis acceleration
• Computational mechanics
• Bio-inspired lattice structures
• Impact resistance

Abstract

Modern structural design and metals processing require fast, reliable predictions across large design and operating spaces, yet brute-force finite element workflows remain too slow for timely screening, inverse design, or process decisions. This study develops conventional deep neural networks and deep operator-learning surrogates that predict full temperature and stress fields from process and geometry inputs, accelerating evaluation by orders of magnitude relative to finite element analysis (FEA). Bio-inspired, low-porosity lattices take design inspiration from sheep horns and hooves and are evaluated under dynamic transverse compression. A gated recurrent unit (GRU) predicts stress-strain curves and energy absorption from geometric inputs. Systematic variations of design parameters, such as tubule shapes and orientations, enable the creation of a dataset comprising 128,000 designs. This approach reveals key design trends and facilitates the enhancement of energy absorption for applications like protective armor. Building on these tools, bio-inspired hierarchical honeycombs (BIHH) are developed by augmenting conventional hexagonal honeycombs with triangular, square, and circular variants, integrating smaller, closed topologies inspired by natural structures such as bamboo, beetles, and crab chelae. These designs, characterized by cross-sections, heights, hierarchy levels, and vertex topologies, are explored using a parameter-based geometry generation framework. A hybrid Transformer-GRU model is trained on geometric features and serves as an efficient surrogate for FEA, enabling rapid exploration of the design space and facilitating impact-resistant design optimization. Inverse design through a genetic algorithm further optimizes BIHHs, identifying novel hierarchical configurations with high energy-absorbing capabilities and stable deformation. For additive manufacturing and materials processing, newly devised deep operator networks (DeepONets), including a sequential, history-aware DeepONet and a geometry-aware ResUNet-DeepONet, predict complete temperature and stress fields under changing loads, histories, parameters, and geometries several orders faster than FEA. In the additive manufacturing case, the ResUNet-DeepONet averages about 0.015 seconds per design versus about 631 seconds for a corresponding FE simulation, roughly 43,000 times faster, with a mean relative absolute error of less than 5 percent across temperature and stress for the median test case. For continuous casting, the sequential DeepONet runs in about 0.018 seconds per case versus about 333 seconds for a FE simulation, about 18,000 times faster, with lower mean absolute errors compared to a GRU model with the same number of parameters. These advances matter at scale. Continuous casting accounts for approximately 95 percent of global steel production, so even small defect reductions translate into large economic and emissions benefits. Fast, generalizable surrogates unlock preliminary evaluation and design optimization that classical solvers cannot deliver on industrial timelines. Together, the results provide a scalable path from expensive simulations to fast, accurate design and process decisions across impact-resistant structures and industrial metal processing.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132763

Copyright and License Information

Copyright 2025 Shashank Kushwaha

Owning Collections