Boosting the Performance of Machine Learning Design Tools with Synthetic Data

Context

Machine learning has emerged as a powerful tool for quickly and cost-effectively evaluating new designs in engineering. By leveraging machine learning, a well-designed user interface can democratize access to science and engineering knowledge, benefiting teams involved in upstream design, manufacturing, and procurement. However, its effectiveness is largely contingent on the quality and breadth of the data used to train the models. This is where synthetic data generation can play a significant role.?

Synthetic Data and Its Role

Synthetic data generation involves creating artificial data that closely mimic the characteristics of real-world data. This is often necessary when actual data is sparse, missing, or imbalanced (too many instances of one class and not enough of another). In engineering, Finite Element Analysis (FEA) simulations can generate synthetic data to address these issues.

In traditional engineering practices, acquiring labelled data for failed designs is often limited due to the rarity of failures and safety considerations. The absence of failure data hampers the ability of machine learning models to learn from critical scenarios and accurately predict potential structural weaknesses. To address this limitation, engineers can employ FEA simulations to generate synthetic failure data. By intentionally introducing faults, varying material properties, or imposing extreme loading conditions, simulated failures can be created, providing a rich dataset for training machine learning models.

Synthetic data generation through FEA simulations offers several advantages. Firstly, it allows engineers to explore a wide range of failure scenarios that may be unlikely to occur naturally. This expands the training dataset and exposes the models to a diverse set of structural behavior. Secondly, synthetic data generation provides an opportunity to examine the failure mechanisms and their correlations with specific design variables, enabling valuable insights into the underlying physics and facilitating improvements in future designs.

Integrating synthetic failure data with physical test-based successful design evaluation data enables machine learning models to learn from both positive and negative outcomes. This balanced dataset provides a comprehensive understanding of the factors contributing to successful designs, as well as the warning signs of potential failures. Learning from a broader spectrum of data empowers machine learning models to generalize better and make accurate predictions, ultimately enhancing the reliability of design evaluation software.

It is important to note that the use of synthetic data does not diminish the importance of real-life failure data. Real-world case studies and failure investigations remain critical for validating and refining machine learning models. The combination of real and synthetic data creates a powerful training framework that bridges the gap between theory and practice, leveraging the strengths of both approaches.

Example : Beauty Cream Tube Structural Design Evaluation

To illustrate this approach, consider the structural evaluation of beauty cream tubes. A software was developed to evaluate tube strength using various tests such as top-down compression, side squeeze, burst test, and seal integrity. Historical data from different designs across various stock keeping units (SKUs) were available, spanning several decades. The first step was to pre-process this data and create labelled tabular data required for machine learning, connecting the data to the pass-fail result of each test. However, there were gaps in the converted legacy information due to missing data for different vendor materials, as well as limited data on failures in the tests since all SKUs were designed to pass these tests. These gaps were addressed by setting up a large Design of Experiments (DoE) using a validated and accurate FEA model covering several cases of synthetic (virtually induced ) failures. This allowed the machine learning calculator to be well rounded with a more reliable and accurate design evaluation software for beauty cream tubes.

To Summarize?

The challenges associated with acquiring failure data for training machine learning-based structural design software can be effectively addressed through synthetic data generation using FEA simulations.?

By intentionally introducing failures and exploring a wider range of scenarios, engineers can fill in missing data, enrich the training dataset, and enhance the accuracy and reliability of machine learning models.?

The integration of synthetic data with real-life failure data offers a balanced perspective, enabling engineers to develop more robust structural designs and drive innovation in the field of engineering.?

Embracing the potential of synthetic data and FEA simulations pushes the boundaries of engineering design, leading to a safer and more efficient future.

#machinelearning #engineeringdesign #syntheticdata #datascience #FEASimulations #DesignEvaluation #artificialintelligence #innovation #dataquality #technologyintegration