FilDeep: Learning Large Deformations of Elastic-Plastic Solids with Multi-Fidelity Data

AI summary

Overview

Research area: Scientific machine learning for computational mechanics — specifically, deep learning surrogate models for the large elastic-plastic deformation of metals, combined with multi-fidelity (MF) data fusion.

Technical level: Advanced. The paper assumes familiarity with continuum mechanics (finite strain theory, elastic-plastic constitutive behavior), attention-based neural architectures (multi-head self/cross-attention, Fourier neural operators, DeepONet), and multi-fidelity learning.

One-sentence scope: The paper proposes FilDeep, a fidelity-based deep learning framework that trains jointly on low-fidelity and high-fidelity simulation data to predict the final shape of metal workpieces in stretch bending, and validates it against single-fidelity and multi-fidelity baselines plus the Finite Element Method.

What This Paper Is About

Simulating large deformations of elastic-plastic solids — such as stretch bending, where a clamped metal workpiece is stretched and bent around a mold and then released — is important for manufacturing but expensive to compute. Deep learning could replace traditional numerical simulation, but it needs large, accurate datasets, and there is a trade-off: high-fidelity (HF) simulations are accurate but yield few samples, while low-fidelity (LF) simulations yield many samples but are biased. The paper's goal is to escape this quantity-accuracy dilemma by training a single model on both fidelity levels at once, rather than on either alone.

Key Contributions

1. Conceptual: The authors identify a general "quantity-accuracy dilemma" in industrial dataset generation, formalized as an Efficiency-Quantity-Accuracy (EQA) principle: high efficiency, large quantity, and precise accuracy cannot be achieved simultaneously in any simulation system. FilDeep is presented as the first DL framework to leverage MF data for large deformation problems of elastic-plastic solids.

2. Technical: An Attention-enabled Cross-Fidelity (ACF) module that captures long-range relationships across MF data for external-condition-aware modeling. The authors state this is the first work to leverage attention mechanisms in MF learning.

3. Dataset: The first-ever MF dataset for large deformation problems, constructed using over 25,000 CPU hours, with all parameters derived from real-world applications, containing 3,000 LF instances and 300 HF instances.

4. Empirical and deployment: Extensive experiments showing state-of-the-art performance versus DL and MF baselines, plus successful deployment in a real-world manufacturing factory, where efficient inference and robust accuracy are reported as competitive.

Main Findings

• Single-fidelity training is the weakest: Across all baselines, models trained on LF data only perform worst, attributed to inherent bias in LF data. Training on HF data only is limited by data scarcity.

• Simply mixing MF data does not help: Naively mixing LF and HF data provides little to no benefit and in some cases causes slight negative effects, because LF and HF outputs differ for the same input and this harms stability and convergence.

• FilDeep achieves the best results with the Transformer backbone: 0.50 mm MAD, 70.83% 3D IoU, and 1.44 mm TE. The authors suggest this comes from synergy between the Transformer's attention design and the ACF module rather than expressive capability alone.

• FilDeep extracts signal from biased LF data: The authors report that FilDeep successfully extracts useful information from noisy and biased LF data, gaining performance by capturing LF-to-HF relationships.

• Ablation confirms the ACF module and residual connection: Against the full FilDeep (0.50 mm MAD / 70.83% 3D IoU / 1.44 mm TE), removing the residual connection degrades to 0.65 mm (+30.00%), 61.88% (−12.64%), 1.84 mm (+27.78%). Residual MFNN reaches 0.75 mm (+50.00%), 55.39% (−21.80%), 2.17 mm (+50.69%). Residual MF-DeepONet reaches 0.68 mm (+36.00%), 59.78% (−15.60%), 2.87 mm (+99.31%). Vanilla MFNN reaches 7.63 mm (+1426%), 12.71% (−82.06%), 19.56 mm (+1258%). Vanilla MF-DeepONet reaches 7.60 mm (+1420%), 13.29% (−81.24%), 19.53 mm (+1256%). The authors attribute this gap to existing MF baselines being designed for simple structured problems relying on feature-level mappings.

• Training protocol for the comparison: Each model was trained on 60 HF samples, with 30 HF samples for validation and 30 for testing. Each baseline was trained separately using single LF data, single HF data, simply mixed MF data, and FilDeep-supported MF data. Within FilDeep, each baseline functions as the LF processor.

• Metrics used: Mean Absolute Distance (MAD), 3-Dimensional Intersection over Union (3D IoU), and Tail Error (TE), plus inference time for efficiency comparisons against FEM. TE targets error accumulation on the tail surface.

Methodology in Plain English

Problem representation. The input is split into a workpiece part and an external-conditions part: x = (w, e) = (s, l, p, m). The workpiece is represented by its 2D cross-section s (as an image, R^{1×H×W}) and a characteristic line l (M points in 3D, R^{M×3}), since the cross-section shape stays constant through the process. External conditions are the loading arm's 6-degree-of-freedom motion parameters p (R^{1×6}) and the mold's characteristic line m (R^{M×3}). The output y (R^{M×3}) is the workpiece's final characteristic line after springback.

Two sub-networks. Following the intuition that the HF output can be modeled as a function of the input and the LF output, the framework approximates y_L = N_L(x; θ_L) and y_H = N_H(x, N_L(x; θ_L); θ_H).

Encoder-processor-decoder architecture. The design mirrors the simulation process. Encoders cover the start-up phase: a Characteristic Line Encoder (CLE) for the workpiece and mold, a Cross-Section Encoder (CSE) using Signed Distance Function representations with a frozen pre-trained ResNet plus trainable convolutional layers, a Motion Parameter Encoder (MPE), and fusion encoders (WRFE, ECFE). Two processors handle LF and HF separately. Decoders map features back to physical space: a Cross-Section Decoder (CSD) built on a Variational Auto Encoder structure with deconvolution and interpolation, and a Characteristic Line Decoder (CLD) that approximately reverses the CLE.

Shared modules and shortcuts for sparse HF data. Because the simulation start-up phase is consistent across fidelity levels, the encoders and decoders are shared. The HF processor also reads the LF processor's output as an input enhancement and adds an LF-to-HF residual connection, so it only needs to learn the gap between fidelities rather than the full HF output from scratch.

Attention for cross-fidelity relations. The HF processor stacks K ACF modules, each using multi-head cross-attention (with external conditions as query and the LF-processed workpiece as key/value), then multi-head self-attention, then a feed-forward network, each with residual connection and layer normalization. The authors motivate attention by noting that clamp motion and mold boundaries exert global effects on the workpiece, not just local ones at contact points.

Losses. A Cross-Section Loss computes MSE between the CSD output and the ground-truth SDF. A Characteristic Line Loss uses a coordinated L2 loss to emphasize coordinate axes with larger orders of magnitude, since the three axes have imbalanced value distributions.

Why This Matters

Impact on research: The paper reframes industrial dataset construction as a trade-off rather than a purely budget-driven problem, and demonstrates that multi-fidelity learning with attention can be a practical route when PDE formulations for a problem are unavailable — making physics-informed methods inapplicable. It also contributes the first MF dataset for large deformation problems, stated to be open-sourced with the code.

Real-world applications (from the paper's framing):

• Metal stretch bending for fabrication, the paper's representative application.
• Manufacturing processes involving large deformations of elastic-plastic solids more broadly.
• Factory-floor deployment, where FilDeep was successfully deployed in a real-world manufacturing factory.
• Comparison against FEM as the incumbent numerical method, on the axis of inference time for practical use.

Industry relevance: The research was conducted with an industrial partner (Fuyao Group) and is explicitly positioned around practical deployment, where the reported property of interest is the combination of efficient inference and robust accuracy under scarce high-fidelity data.

Future Directions

• Extending beyond stretch bending: The framework is developed and evaluated on one representative large deformation process; whether the representation scheme (cross-section, characteristic line, motion parameters, mold line) transfers to other forming processes is not established in the paper.
• Scaling and generalizing the MF dataset: The released dataset contains 3,000 LF and 300 HF instances. Whether the multi-fidelity advantage holds at other LF-to-HF ratios, or with more than two fidelity levels, is not reported.
• Replacing the attention backbone or LF processor: The paper shows FilDeep's performance varies depending on the backbone used as the LF processor (Transformer performing best); the design space of LF processors and ACF variants remains open.
• Understanding when data mixing fails: The paper observes that simply mixing LF and HF data can slightly harm performance, and attributes this to differing outputs for the same input; a principled account of when mixed data helps or hurts is not fully resolved in the content presented.

Target Audience

Researchers and practitioners working at the intersection of scientific machine learning, surrogate modeling, and computational mechanics — particularly those applying multi-fidelity methods to physical simulation where explicit PDEs are unavailable. It is also relevant to engineers evaluating deep learning against FEM in manufacturing settings, and to readers interested in attention-based architectures for non-standard, unstructured physical problems. The paper's technical density (attention mechanisms, finite strain mechanics, multi-fidelity learning) makes it most accessible to readers already comfortable with at least one of those areas.

Authors’ abstract

The scientific computation of large deformations in elastic-plastic solids is crucial in various manufacturing applications. Traditional numerical methods exhibit several inherent limitations, prompting Deep Learning (DL) as a promising alternative. The effectiveness of current DL techniques typically depends on the availability of high-quantity and high-accuracy datasets, which are yet difficult to obtain in large deformation problems. During the dataset construction process, a dilemma stands between data quantity and data accuracy, leading to suboptimal performance in the DL models. To address this challenge, we focus on a representative application of large deformations, the stretch bending problem, and propose FilDeep, a Fidelity-based Deep Learning framework for large Deformation of elastic-plastic solids. Our FilDeep aims to resolve the quantity-accuracy dilemma by simultaneously training with both low-fidelity and high-fidelity data, where the former provides greater quantity but lower accuracy, while the latter offers higher accuracy but in less quantity. In FilDeep, we provide meticulous designs for the practical large deformation problem. Particularly, we propose attention-enabled cross-fidelity modules to effectively capture long-range physical interactions across MF data. To the best of our knowledge, our FilDeep presents the first DL framework for large deformation problems using MF data. Extensive experiments demonstrate that our FilDeep consistently achieves state-of-the-art performance and can be efficiently deployed in manufacturing.