AI is revolutionizing engineering design and scientific research across different business segments and research fields with speed-ups and accuracy improvements as stated clearly by Christopher Bishop, head of AI4Science at Microsoft Research: “Over the last year or two, we have seen the emergence of a new way to exploit deep learning […] the data that is used to train the neural networks itself comes from numerical solution of the fundamental equations of science rather than from empirical observation. [..] Once trained, the emulator can perform new calculations with high efficiency, achieving significant improvements in speed, sometimes by several orders of magnitude.” 

Notable examples of this fruitful intersection come from nearly all fields where there is a large use of complex simulation tools, such as the development of new molecules in the pharmaceutical or semi-conductor industries (doi.org/10.3389/frhem.2024.1305741), the development of advanced control algorithms for fusion reactor (doi.org/10.1038/s41586-021-04301-9) or the prediction of all the possible scenarios for anticipating extreme weather events (arxiv.org/abs/2306.03838).

AI can add value to engineering design right now, revolutionizing many applications, such as for example:

• Efficient data-matching and reconciliation of simulation models
• Faster engineering design optimization (parametric / parameter-free)
• Accelerated sensitivity and reliability analysis
• Increased accuracy of lower fidelity numerical simulations thanks to data-driven models trained on higher fidelity data

Within Baker Hughes, we have been developing a series of successful initiatives in the past years that show the benefit of having AI specialists work alongside design engineers. From this experience, we learned significant lessons and we would like to share some of them in the form of example projects.

Monte Carlo based Reliability Calculation

The Monte-Carlo simulation strategy is a common approach for reliability analysis since it allows to estimate the probability of extreme events as well. The number of simulation runs however increases with the rarity of the extreme event, therefore even if the simulation tool is efficient and parallelizable, the time requirements for each scenario are in the order of magnitude of several hours (>30). The use of faster surrogate models could be a viable solution, but the need to predict extremely rare failure events, only found in the tails of the data distribution, poses a challenge to training an effective model. We have therefore put in place a hybrid two-step approach, with a first classifier model that predicts the risk of failure of the given sample, and based on its prediction, either the real simulator is ran, or the surrogate model result is trusted. After an initial training with a limited and optimized dataset (based on active learning and space filling techniques), the hybrid approach allows to run only the most significant simulations. The obtained results are consistent within a 0.01% error with the original process, but requiring only 10% of the computing time. This enabled a step change in the design process, because design engineers now can test many more multiple design alternatives and choose the optimal one in terms of reliability, in the same amount of time.

Design optimization using surrogate models of CFD simulations

Engineering designs are usually described as parametric CAD models. Predefined parametrizations, however, have some limitations: they may not be the most informative set of features for surrogate models. They cannot describe similar geometries, generated with different parametrizations. Finally, manufacturing variations don’t always map directly to design parametrization. The solution is to let AI learn a latent parametrization from a dataset of geometries. We have worked on a test case for the design optimization of a centrifugal compressor stage for which different operating conditions and different geometries had been already simulated. Considering the limits of the dataset (small number of samples and limited geometric variations) we tested various mesh encoding techniques (Autoencoders, PCA) to derive the latent parametrization of the geometries. We combined the learned coefficients with the operating conditions and the performance KPI, to generate the input data for training the machine algorithms to predict compressor performance curves. Besides the improved performance of the surrogate model when compared to others using the original CAD parametrizations as input, the additional advantage of this technique is the possibility to use the learned representation (capable of reconstructing the geometry itself) for generating a more efficient parameter space for the subsequent optimization steps.

Efficient data acquisition strategy for building surrogate model of FEM simulations.

Within the framework of design exploration, we have encountered many times the task of building reduced order surrogate models of complex simulations, to have a faster tool for parametric design exploration and multi-objective optimization. Usual model development follows a two-step approach: first, a space filling Design of Experiment (DoE) technique is employed to generate the dataset (containing the desired output from the post-processing of the simulations), then the surrogate models are trained and validated using the available data points. We empirically demonstrated that a continuous training and validation strategy during data acquisition (active learning) significantly reduces the computational time without affecting accuracy. The reason is that it stops acquiring new data points (time-consuming simulations) once the model performance stops improving. Active learning however struggles when the machine learning model needs to predict multiple targets at the same time. By combining active learning with iterative space filling techniques and parallel processing, we were able to reduce computational time of about 50%.

Future of AI for Design in the industry

The application of advanced models that will replace detailed simulation tools is expected to grow constantly, but the need for training data is expected to remain the main bottleneck in the adoption of these tools, unless an integrated technological workflow is put in place. We see that some software vendors are already working on this direction, combining Machine Learning and traditional simulators in the same design process, to harness the strengths of both approaches.

Breakthrough will come from the usability of simulation tools with lower adoption barriers and speedups given the introduction of AI-based assistants that can communicate in natural language with design and simulation engineers. These models will allow to directly interact with design practices, or to perform changes automatically and safely thanks to more automatic control on the cases setups and assisted pre-processing steps. For example, geometry preparation, assemblies and mesh generation phases are expected to be more and more automatized thanks to AI assistants.