Advancements in synthetic data extraction for industrial injection molding

AI summary

Overview

Research area: Industrial machine learning, specifically synthetic data generation and time-series classification for injection molding quality control.

Technical level: Intermediate. The paper assumes familiarity with neural networks, LSTMs, and standard classification metrics, but its core argument about mixing simulated and real data is accessible to a general technical reader.

Scope: A single-study investigation into whether adding simulated injection molding cycles (up to 30 percent of the training set) to a small real dataset changes the accuracy, robustness, and generalization of an existing LSTM quality classifier, conducted at Rosenheim Technical University of Applied Sciences (arXiv:2511.08117v1 [cs.AI], 11 Nov 2025).

What This Paper Is About

Machine learning can optimize industrial production, but collecting and labeling real data is slow and expensive, and injection molding machines mostly produce near-identical parts, so gathering genuinely varied or defective examples means deliberately wasting material, machine hours, and skilled labor. This paper asks whether synthetic production cycles, created in CAD and a production simulator, can be blended into a small real training set without damaging the model. The goal is to find a proportion of synthetic data that improves robustness while preserving the authenticity of the real data.

Key Contributions

1. A four-stage pipeline for industrial synthetic data: simulating production cycles, labeling the simulated cycles, enriching training sets with them, and evaluating on real-only validation data.
2. A systematic sweep of synthetic data proportions from 0% to 30% in 5% increments (7 training sessions of 50 runs each) applied to an existing LSTM quality classifier.
3. A second experiment in which the total training set size is held constant while real data is progressively replaced by synthetic data, isolating the effect of the real-to-synthetic ratio.
4. Evidence that a deliberately over-representation of defective cycles in the simulation can compensate for the class imbalance typical of real production data.

Main Findings

• Costs of the fully real baseline: With 0% synthetic data, average validation accuracy was 93.6% and training accuracy 92.8%. In the detailed table these appear as validation accuracy 0.9363, validation loss 0.05935, training accuracy 0.9284, training loss 0.06972, F1 score 0.9238, and AUC-ROC 0.9743.

• Synthetic data reduced training accuracy but left validation accuracy roughly stable: At 30% synthetic data the average validation accuracy dropped slightly to 92.3% (0.9233 in the table, validation loss 0.07094) while training accuracy fell to 85.5% (0.8523, training loss 0.1106). The paper reads this as the model becoming harder to overfit rather than becoming less accurate.

• Predictive metrics held up: F1 score was 0.9238 at 0% synthetic data and 0.908 at 30%; AUC-ROC was 0.9743 at 0% and 0.9677 at 30%. The paper calls these changes negligible.

• A mid-range dip is visible in the results: Validation accuracy for the full sweep was 0.9363 (0%), 0.9164 (5%), 0.9221 (10%), 0.9007 (15%), 0.9165 (20%), 0.9139 (25%), and 0.9233 (30%); F1 scores were 0.9238, 0.8979, 0.907, 0.8784, 0.9009, 0.8945, and 0.908 respectively.

• Constant-size training sets: When the overall data set size was fixed and the real portion shrank from 100% (737 real cycles, 0% synthetic) to 55.2% (407 real cycles, 44.8% synthetic), validation accuracy dropped from 0.9363 to 0.8273 and training accuracy from 0.9284 to 0.82. The paper reports that with more than 15% synthetic data the F1 score improved relative to the absence of synthetic data in Table 3, and that a modest dataset supplemented with synthetic data surpassed a non-augmented one on validation accuracy.

• Robustness under perturbation: Comparing 0% and 30% synthetic data runs, validation accuracy and loss curves stayed relatively similar, while the training accuracy and loss curves scattered more for the 30% synthetic dataset.

• Data asymmetry: The real data comprised 275 injection molding processes, quadrupled to 1100 labeled cycles split into 56.5% good and 43.5% not good. The synthetic set comprised 100 cycles, quadrupled to 400, split into 60% not good and 40% good, intentionally favoring defects.

• Stated limitation: The LSTM reached an average validation accuracy of about 94% on the real data set, but the authors say this cannot be generalized because the real dataset is too small, and synthetic labeling had to be done manually.

Methodology in Plain English

The researchers first recorded real production: 275 injection molding cycles, with an experienced injection molder judging part quality for each one. Key process parameters such as piston stroke, rounds per minute in odization, injection volume, back pressure, and holding pressure were varied on purpose to create a range of conditions. Because sensor samples came at a fixed 10 ms interval, each cycle was enlarged fourfold by storing the first through fourth step at 10 ms intervals in separate files, yielding 1100 labeled cycles. One third of the real data, 363 datasets, was held out for validation, leaving 737 for training.

Synthetic data came from two sources working together: a CAD program that simulated the plastic filling study (material flow, cooling dynamics, pressure distribution) and a production simulator that is a clone of a real machine control, accepting settings such as injection speed, changeover point, and holding pressure level and outputting process parameters such as screw position, residual compound cushion, and ejector speed. The authors note the simulator's own values are not close to reality, which is why the CAD program is used alongside it. 100 synthetic cycles were generated and quadrupled to 400, and their quality labels were assigned manually using knowledge of the simulator parameters plus the CAD fill study output.

An existing LSTM classifier was then trained on mixtures of these two pools. The network takes 34 input features, produces one output, and has three LSTM layers of 100 units each, with dropout layers after each (dropout 0.1598 after the first two, 0.279 after the last) and the first two layers configured with return_sequences set to True. Training used the Adam optimizer, MSE loss, a batch size of 64, 50 epochs, and a learning rate of 0.0001175. The authors cite "Sequence to Sequence Learning with Neural Networks" as the architectural inspiration and describe the setup as a unidirectional many-to-one configuration; they state they did not extensively compare other LSTM architectures, choosing the best performer from an initial round of experimentation. All computation ran on a 12th Gen Intel Core i9-12900KS 3.40 GHz CPU with 64GB memory and an NVIDIA GeForce RTX 3090 Ti GPU, using Tensorflow 2.11.0 and Keras 2.11.0.

Why This Matters

For research, the paper offers a concrete, reproducible demonstration that simulated industrial time-series data can be blended into small real datasets without collapsing model performance, and it provides a full accuracy/loss/F1/AUC-ROC record across seven synthetic fractions rather than a single headline number.

Real-world applications suggested by or implied in the paper:

• Reducing manual labor, machine usage, and material waste during injection molding process optimization.
• Improving defect detection and quality classification when real production data is small and unbalanced, since defective parts are rare in normal production.
• Serving companies that lack the resources to invest in extensive data collection and maintenance.
• Generating and labeling production errors in a simulated environment instead of producing broken parts on an actual machine.

Industry relevance: injection molding is a high-volume manufacturing process where every experimental production run costs machine hours and material. A method that shifts experimentation into simulation addresses the industry-wide bottleneck the paper identifies, namely that data collection, cleaning, analysis, visualization, and feature engineering consume most of the time in an ML project.

Future Directions

1. Increase the synthetic fraction in the data as much as possible, extending beyond the 30% ceiling tested here.
2. Train a Generative Adversarial Network to produce synthetic data that also accounts for the noise and uncertainties of real-world data.
3. Train an independent classifier to label synthetic production cycles automatically, replacing the current manual labeling and freeing time for data augmentation.
4. Train a GAN that generates data without simulation, with the aim of producing more realistic, high-quality synthetic data, and validate against larger, more representative real-world datasets.

Target Audience

Industrial ML practitioners and manufacturing engineers who need to train models on scarce production data; researchers working on synthetic data generation, sim-to-real transfer, or time-series classification; and quality assurance teams in plastics and injection molding who are weighing whether simulated cycles can substitute for costly real-world experimentation.

Authors’ abstract

Machine learning has significant potential for optimizing various industrial processes. However, data acquisition remains a major challenge as it is both time-consuming and costly. Synthetic data offers a promising solution to augment insufficient data sets and improve the robustness of machine learning models. In this paper, we investigate the feasibility of incorporating synthetic data into the training process of the injection molding process using an existing Long Short-Term Memory architecture. Our approach is to generate synthetic data by simulating production cycles and incorporating them into the training data set. Through iterative experimentation with different proportions of synthetic data, we attempt to find an optimal balance that maximizes the benefits of synthetic data while preserving the authenticity and relevance of real data. Our results suggest that the inclusion of synthetic data improves the model's ability to handle different scenarios, with potential practical industrial applications to reduce manual labor, machine use, and material waste. This approach provides a valuable alternative for situations where extensive data collection and maintenance has been impractical or costly and thus could contribute to more efficient manufacturing processes in the future.