Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test
Zhengni Yang 1 , 2 , 3 , Rui Yang 1 , 1 , Weijian Han 2 , 1 , Qixin Liu 1 · 2025
Read Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test by Zhengni Yang 1 , 2 , 3 , Rui Yang 1 , 1 , Weijian Han 2 , 1 , Qixin Liu 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-17680.txt.
Introduction
Zhengni Yang, Rui Yang, Weijian Han, and Qixin Liu present a deep learning method for materials tests, in two thousand twenty-five. A small punch test can probe strong steel with a tiny sample. The hard part is turning its load and movement data into a full stress and strain curve.
The team first turns the data series into images that hold both space and time patterns. A sequence model with a focus step then reads those images and predicts the curve. In tests, the errors stayed low across the steels tried. The authors offer the method as a faster, more exact stand-in for older lab routes.
Catalog record
| CREATOR | Zhengni Yang 1 , 2 , 3 , Rui Yang 1 , 1 , Weijian Han 2 , 1 , Qixin Liu 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 21 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-17680
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test ∗ Corresponding author: Rui Yang ([email protected]), Weijian Han ([email protected]) This research is partially supported by: Jiangsu Provincial Qinglan Project and Suzhou Science and Technology Programme (SYG202106). Zhengni Yang 1 , 2 , 3 , Rui Yang 1 , 1 , Weijian Han 2 , 1 , Qixin Liu 1 Abstract This paper introduces a novel deep-learning approach to predict true stress-strain curves of high-strength steels from small punch test (SPT) load-displacement data. The proposed approach uses Gramian Angular Field (GAF) to transform load-displacement sequences into images, capturing spatial-temporal features and employs a Sequence-to-Sequence (Seq2Seq) model with an LSTM-based encoder-decoder architecture, enhanced by multi-head cross-attention to improved accuracy. Experimental results demonstrate that the proposed approach achieves superior prediction accuracy, with minimum and maximum mean absolute errors of 0.15 MPa and 5.58 MPa, respectively. The proposed method offers a promising alternative to traditional experimental techniques in materials science, enhancing the accuracy and efficiency of true stress-strain relationship predictions. I Introduction Accurately determining the true stress-strain relationship of materials is a fundamental aspect of engineering design and structural integrity assessments [ 23 , 21 ] . Traditional methods such as tensile, compression and shear tests are costly, time-consuming, and sample-intensive [ 10 , 11 ] . The small punch test (SPT) offers a viable alternative, reducing sample sizes and minimizing component damage, particularly in challenging testing scenarios [ 1 , 12 , 3 , 9 ] . Advancements in machine learning have transformed the fields of materials engineering and fault diagnosis, offering precise predictive capabilities [ 20 , 19 , 2 , 24 ] , and the machine learning algorithms accelerate material characterization and decrease reliance on traditional methods by capturing complex data relationships [ 8 , 15 , 22 , 17 ] . Among the various machine learning methods, the sequence-to-sequence (Seq2Seq) model has attracted considerable attention for its effectiveness in handling sequential data [ 16 ] . The Seq2Seq model was initially designed for natural language tasks due to its effectiveness in handling sequential data through an encoder-decoder architecture, which predicts and decodes sequences [ 6 ] . The Seq2Seq model captures intricate dependencies in complex data like load-displacement and stress-strain sequences, and the encoder-decoder structure transforms input data into a context vector for accurate output generation [ 4 ] . This ability to precisely handle sequential data highlights the potential of the Seq2Seq model as a robust tool for modeling and predicting material behaviors, offering an efficient alternative to traditional experimental approaches in materials science. In material science, the relationship between SPT data and stress-strain data can be viewed as a translation task, akin to the mapping between source and target languages. In this study, the Seq2Seq model is applied to predict stress-strain curves from SPT data, using the load-displacement curve as the input sequence and the corresponding stress-strain curve as the output sequence. This application allows the model to learn the complex mapping between the input and output sequences, offering a novel approach for predicting stress-strain sequence data from SPT data. The Seq2Seq model provides an efficient and accurate means of predicting material behavior, reducing the reliance on extensive experimental data and offering new possibilities for rapid material property evaluation. Integrating the Seq2Seq model into materials science highlights the transformative potential of machine learning and deep learning in predicting and understanding material behaviors. By using SPT data, the proposed approach offers a more efficient and cost-effective alternative to traditional methods, accelerating material characterization. This paper presents a novel method for predicting the true stress-strain relationship of high-strength steels using the Seq2Seq structure, advancing the application of machine learning in material behavior prediction. The main contributions of this study are as follows: 1. This study introduces a novel deep-learning framework that employs a Seq2Seq model with multi-head cross-attention to improve prediction accuracy and efficiency compared to traditional methods; 2. This study proposes a feature extraction method using multi-scale convolution layers to capture features from the original sequence and Gramian Angular Field (GAF) images, further enhancing the model’s predictive accuracy. The remainder of this paper is organized as follows: Section II details the proposed methodology, Section III presents the experimental results and discussions, and Section IV concludes the study and suggests directions for future research. II Methodology This study proposes a novel method to predict true stress-strain curves from SPT data. The proposed approach utilizes SPT data as input and employs a Seq2Seq model combined with a multi-head cross-attention mechanism to accurately predict stress-strain curves for high-strength steel materials. The architecture of the proposed method is illustrated in Fig. 1 . This section provides a detailed description of the data generation process, the feature extraction algorithms, and the structure of the proposed model employed in this approach. Fig. 1: Structure of Proposed Approach II-A Data Generation This study generated training and testing data for high-strength steel using LS-DYNA software to simulate load-displacement and stress-strain curves. The finite element method (FEM) modeled the SPT and tensile tests, accurately replicating sample geometry and loading conditions. Simulations covered materials with varying properties and thicknesses from 1 to 4 mm, using Young’s modulus of 70,000 MPa and a Poisson’s ratio of 0.35. For the SPT simulation, the load was set to 20 kN, and the puncher was pressed into the specimen at a constant displacement rate of 10 mm/min, with load and displacement data recorded throughout the process. For tensile testing, the stretching rate was set to 7.2 mm/min, and stress-strain data were collected. This setup enabled the generation of 5,000 high-strength steel samples with varying mechanical behaviors, forming the basis for the training and testing datasets used in this study. Fig. 2 illustrates the data generation process using LS-DYNA: Fig. 2(a) presents a schematic of the simulation mechanism, showing the stages of punch impact on the model, while the Fig. 2(b) depicts the simulation setup for tensile testing. Overall, in this study, 5000 high-strength steel samples with varying mechanical behaviors are generated for training and testing. (a) Progressive Stages of Punch Impact on the Simulation Model (b) Progressive Stages of Tensile Test on the Simulation Model Fig. 2: FEM Configurations for Small Punch Test and Tensile Test II-B Feature Extraction Algorithm 1 Training Procedure for the Proposed Model 1: Input: Load-displacement data \mathcal{D} 2: Output: Predicted stress-strain data \mathcal{P} 3: Preprocess Data: 4: Compute Gramian Angular Field \mathcal{G} using ( 1 ) and ( 2 ) 5: Extract 2D features \mathcal{F}_{2D} from \mathcal{G} using 2D Conv layers 6: Extract 1D features \mathcal{F}_{1D} from \mathcal{D} using 1D Conv layers 7: Obtain feature matrix \mathcal{M} using ( 3 ) 8: Train Seq2Seq Model: 9: Initialize encoder f_{\text{encoder}} , decoder f_{\text{decoder}} , and cross-attention \mathcal{A} 10: Set optimizer to Adam and loss function 11: for each epoch \epsilon=1,\dots,\text{num\_epochs} do 12: for each batch \mathcal{B} in \mathcal{M} do 13: Prepare input sequence \mathcal{X} and target sequence \mathcal{Y} 14: Encode \mathcal{X} to obtain hidden states \mathcal{H} using ( 4 ) 15: Initialize batch loss \mathcal{L}\leftarrow 0 16: for each time step t=1,\dots,\text{length of }\mathcal{Y} do 17: Generate decoder output \mathcal{O}_{t} using ( 5 ) 18: Compute context vector \mathcal{C}_{t} using ( 6 ) 19: Compute prediction \mathcal{P}_{t}