According to Gartner, companies that have adopted AI engineering practices to develop and manage adaptive AI systems have a clear competitive advantage: by 2026, such pioneers will outpace their competitors by at least 25 percent in terms of the number and time required to operationalize AI models. In order to stay ahead of the game, companies must continue to drive the introduction of AI and develop new use cases for themselves. The article from MathWorks explains which AI trends engineers should keep an eye on in 2023 and which challenges need to be overcome:

Physics-based AI - models with principles and rules from the real world

AI is spreading into more and more areas of research, for example in complex technical systems. The combination of data and physics via neural ODEs (ordinary differential equations) or PINNS (physics-informed neural networks) has great potential. Simulations are at the heart of physics-based artificial intelligence: Complex models can be configured as variants within a simulation and offer developers a quick switch between models to obtain the most accurate solutions. Reduced order modeling (ROM) with physically based reduction models is also an important trend. By using AI, simulations can be accelerated by replacing an extremely computationally intensive first-principles model of a system - while maintaining accuracy.

Collaboration - Free access to AI is spreading

Researchers, engineers and data scientists should continue to expand their cross-functional and cross-industry collaboration to think of innovative solutions from different angles. To make the latest models available on demand and enable users to build on the latest research results in the shortest possible time, network-based version management services for software development projects such as GitHub are a good option. Open source solutions are also becoming increasingly popular, as engineering teams often work with models from different frameworks. Greater networking between science, academic research institutions and companies is driving AI research into topics such as physics-based machine learning and biomedical image processing.

Companies focus on smaller, easier-to-explain AI models

AI users are increasingly realizing that they need to provide, adapt and explain models in order for these models to be relevant. The explainability of models and corresponding applications are therefore moving into the focus of engineers. To meet the requirements for low-cost, low-power devices with explainable outputs, engineers are turning to traditional machine learning models and parametric models. These are compact, have low memory requirements and are easier to interpret. When newer, more memory-intensive models are needed, quantization and pruning techniques offer ways to compress the models. So, if needed, engineering teams can leverage interpretability, quantization and pruning to extend the use of AI, including deep learning and traditional machine learning models, to traditional model development.

AI is becoming crucial for the design, development and operation of modern technical systems

AI is becoming more and more prevalent in all industries and applications and will be crucial for technical progress and the development and operation of modern technical systems in the future. In more established fields of activity where AI has just been introduced, engineers need background information and specific reference examples to integrate AI into their work. Based on proven examples, engineering teams can contribute data and know-how and integrate AI specifically adapted to their tasks.

What challenges await AI engineers

As different teams are usually responsible for the creation and implementation of AI models, complex challenges arise in the AI environment that engineers need to overcome. The selection of pre-processing algorithms and model training often fall within the remit of data scientists, who focus on accuracy and robustness. However, engineers must also take other criteria into account for successful porting to the target platform. Early testing of algorithms for a feasibility assessment using PIL (Processor-in-the-Loop) can prevent already trained and powerful models from having to be discarded. The training of the AI is also usually implemented in a different programming language than the implementation in the hardware. However, models from the training environment cannot simply be executed on the target hardware without further ado. To overcome barriers between scripting languages, there are runtime interpreters, machine learning compiler frameworks such as Apache TVM or automatic code generation in MATLAB/Simulink.

Finally, the validation of AI models remains an important topic: while AI models are allowed to make mistakes in the training environment in order to learn and improve, errors after implementation on the hardware can lead to major damage in real existing systems. The question of reliable, objectively verifiable criteria for a model to be considered safe will remain an important area of research in the future.

You can find out more about AI modeling, workflows and systems here.

The author: Johanna Pingel is Product Marketing Manager for AI, Mathworks