Leveraging know-how with AI into efficient product design

Artificial Intelligence (AI) involves computers and machines that can perform tasks commonly associated with humans, such as learning, reasoning, and problem solving. While core AI technologies (including neural networks (NN) and machine learning (ML)) have been known for a long time, they have vigorously developed in the past decade, bringing AI to the forefront of our businesses and lives^1,2.

We’re all familiar with increasing AI technology, for example, guiding our commute with real-time navigation and providing personalized content from streaming recommendations to social media feeds. We rely on facial recognition for identification and consult digital voice assistants that become increasingly helpful. AI furthermore influences our lives in less visible areas, such as managing and optimizing energy grids and supply chains.

AI is also finding its way into the manufacturing industry, enhancing efficiency and providing innovation opportunities, up to disruption of industrial practices. Examples include augmenting computer-aided engineering (CAE) with AI/ML into efficient product design³, generative engineering⁴ and achieving high levels of automation and efficiency with intelligent robotics and optimized production lines⁵.

1. Challenge: AI requires high quality data

The phrase “As a stream reflects its source” describes how an output is determined by the characteristics of its origin or inputs. For crude oil, this means that the type of oil is influenced by the environment and organic matter from which it formed. Similarly, for AI models, the predicted output is directly determined by the quality and fidelity of its input (the source data) that is used for training the AI model. Hence, the phrase becomes “An AI model reflects its data.“

When adopting AI in a modelling and simulation context, the key challenge is the need for enough high-quality data, diverse and unbiased, that accurately represents real-world behavior. Such source data can be used to train an AI model, that can subsequently predict real-world behavior in an accurate and efficient way. The data quality has a strong impact on the prediction performance of the AI model. In data science, the phrase “Garbage In, Garbage Out” was coined by IBM programmer George Fuechsel in 1962⁶, indicating that if we put bad information into our computer models, we will get bad predictions out of them. This equally applies to AI: one must avoid that bad input quality, through model training, results in inaccurate / biased and/or noisy predictions of the AI model.

A paradox is that while businesses rely increasingly on data, there’s a tendency to trust the data less. Salesforce⁷ published that 76% of business leaders believe that AI has made a data-driven approach more important than ever. In contrast, only 36% of these leaders trust the accuracy of their company’s data (a drop of 27% in only one year).

To address the above challenge and paradox, we need to provide high quality data as input to the AI model training phase, such that the resulting AI model can accurately predict real-world behavior. The data can come from experimental testing or from simulating high-fidelity models of representative design variants. Data scientists must perform proper data collection, labeling and storage, and ensure that the data is free from inherent biases that could lead to unfair or inaccurate AI models.

Unfortunately, at the start of each new product design challenge, there’s data scarcity or the ‘cold start’ problem in AI: no high-fidelity models / data are available yet for the new design that is envisaged, hence at the start of the project, it’s a large challenge to find relevant data in sufficient quantity.

2. Solution: Exploit data similarity and transfer learning

Over time, most industrial product manufacturers generate a significant number of models, data, and results.  Usually this goes beyond raw data, and includes contextualized, validated, and often labeled knowledge that could be suitable for training AI models. For example, an automotive company performing time-consuming flow-simulations for each of its new designs may have a significant archive of data from its previous designs. Such data sets hold great promises to overcome the initial data scarcity challenge, if the manufacturer can re-use the previous data and know-how (for example obtained from predecessor designs), to train the initial AI model (for the new design). Two steps are taken to achieve this:

1. First, a data similarity metric is adopted to select the most similar past designs (according to the initial view / drawings of the new design).
2. Transfer learning methods⁸ can enable AI to re-use data from previous designs, and adapt that knowledge to the new design, reducing the requirement for new data.

As partner in the ITEA4/VLAIO research project SmartEM, Siemens Digital Industries Software researched and developed new methods for similarity assessment and transfer learning, which enables us to do more with past data in the industrial design engineering process. These new methods contribute to the SmartEM objective to enable the reuse, exchange and integration of computational engineering models, reducing the need for costly design corrections and promoting early data and model exchanges. Figure 2 visualizes the first and most critical step in this process: assessing a quantitative similarity between any pair of geometries in an efficient and accurate way. The methods developed in SmartEM enable speed-ups of > 100.000x without loss in accuracy and with no requirements for any form of training. Such methodologies enable efficient exploration of large historical datasets, highlighting best candidates for transfer learning and leveraging existing knowledge.

The new methods allow us to reduce the reliance on more expensive data generation methods (such as physical testing or high-fidelity model creation and computation), as re-using past data will reduce the computational effort and cost for the new design. By leveraging high-accuracy 3D simulation model results from predecessors⁹,¹⁰ for the training of an initial AI model, the new methods shorten the time to create an AI model that provides reliable predictions of the new product design. This can then be used as a basis to optimize the new design in Simcenter HEEDS¹¹. Simcenter HEEDS is a powerful workflow automation, design space exploration and optimization software package, which can be used as an orchestrator¹² for the predecessor data re-use, and for performing any new calculations, which can be based on AI (for efficient optimization of the new design) and/or on high-fidelity models (e.g. for some new data generation and for validation of the optimum), to discover better designs, faster.

3. Conclusions

Depending on their internal processes and data management efforts, industrial product manufacturers sit on a potential ‘goldmine’ of past models, data and results. The challenge is to tap into past data and leverage it to train an accurate and efficient AI model that can kick-start the new product design cycle. This requires adopting best practices in data management and labeling and requires new methods for design similarity and transfer learning, as presented in this blog post.

Acknowledgements

This research work has been carried out in the frame of the ITEA4 research project 22009 “SmartEM – Open reference architecture for engineering model spaces”, coordinated by Philips Consumer Lifestyle BV, The Netherlands, and with Siemens Industry Software NV (SISW) and KU Leuven as partners in Flanders, Belgium. We gratefully acknowledge ITEA4 and VLAIO (for the Flemish subproject HBC.2023.0520) for their support.

References

1. Pekka Salokannel, The Impact of AI: How AI is Transforming Society, 3D Bear, March 2023. ↩︎
2. Thomas Hartung (ed.), Frontiers in Artificial Intelligence, Frontiers Media, Retrieved 2025. ↩︎
3. Daniel Berger, Justin Hodges, Dirk Hartmann, Beyond traditional engineering simulations: How Machine Learning is pioneering change in engineering, The Art of the Possible, October 10, 2023. ↩︎
4. Robin Bornoff, Dirk Hartmann, Generative Engineering and the role of simulation, The Art of the Possible, November 10, 2023. ↩︎
5. Shirley Segey, AI in manufacturing: Transforming engineering, production and supply chains, Tecnomatix Blogpost, February 27, 2025. ↩︎
6. David Donoho, 50 years of Data Science, Journal of Computational and Graphical Statistics, 26(4), pp. 747-766, 2017. ↩︎
7. Salesforce, The State of Data and Analytics Report, 2024. ↩︎
8. K.J. Voogd, J.P. Allamaa, J. Alonso-Mora, T.D. Son, Reinforcement Learning from Simulation to Real World Autonomous Driving using Digital Twin, arXiv:2211.14874v1 [cs.RO], November 27, 2022. ↩︎
9. Siemens Digital Industries Software, Simcenter 3D, Retrieved 2026. ↩︎
10. Siemens Digital Industries Software, Simcenter Nastran, Retrieved 2026. ↩︎
11. Siemens Digital Industries Software, Simcenter HEEDS, Retrieved 2026. ↩︎
12. Mathijs Vivet, Adrien Scheuer, Stijn Donders, AI-accelerated gear stress analysis, Simcenter blogpost, August 27, 2025. ↩︎