Building a Hybrid AI model in a rigorous engineering world
We all know that AI has the potential to make significant improvements to many industries, by making existing processes and products more efficient, as well as enabling whole new classes of applications.
However, in the world of engineering and science, those benefits are harder to realize, and for several reasons. First, because operating in this industry requires high degrees of precision, validation and traceability – traits that naïve AI is not known for. Secondly, engineering data is diverse and interconnected – not just text, but diagrams, CAD files, molecular descriptors and many more, and our solutions must be consistent across these modalities. Finally, engineers work in a diverse architectural ecosystem, physically deploying solutions onto many different sizes and scales of platform, often with challenging real-world constraints.
Hybrid AI is a new architectural weapon in the Augmented Engineer’s arsenal which helps us unleash the power of AI into the engineers’ world, without compromising on their core beliefs and convictions.
A Hybrid AI system combines AI models that derive their power from the powerful but unreliable statistical modeling of large data sets, with more traditional approaches to reasoning about structured knowledge, to produce hybrid systems that blend the best of both worlds. For example, an AI system for driving an autonomous vehicle needs to be able to do more than just spot road signs and other vehicles – it needs to be able to reason about the physics of how objects move or how human behavior changes in different cultures. Similarly, a generative design system for making aerospace parts will fail if it does not also consider the physical and structural properties of that component.
The practical reality of implementing these multi-faceted Hybrid AI systems is itself a multi-faceted challenge. Teams with expertise in contemporary machine learning methods need to come together with teams experienced in more traditional techniques for representing and reasoning about structured knowledge. It should be no surprise that both sides of this dichotomy are entirely reliant on data. While the Big Data era made claims that “data is the new oil”, the AI era has made that claim a reality, where those in possession of enough high-quality, well-curated data are able to convert that into a new tool with inherent value – an AI model.
In engineering, the data landscape is inherently multi-modal, where data is represented in different forms, for example, a written description of a process and a diagram of that process, and it is vital that consistency between those modalities is maintained. While for many, large language models have become a default paradigm, in this world, that narrow single-modality approach could be a limiting constraint.
Once the data is ready, training or fine-tuning a model can commence, but careful consideration needs to be given to the choices made here. Many different types of model are available, which have radically different capabilities, operational profiles and resource requirements. Choosing the wrong model could lead to massive operational costs, poor accuracy, or skeptical users. Larger models can also have significant environmental impact.
Hybrid AI again comes to the rescue here, allowing us to produce solutions that are much leaner. Why? Because the knowledge-driven part of the model can provide concise rules for how to behave when the pattern-based part of the model doesn’t have any similar prior experience. Purely pattern-driven models need to “see everything” in order to guarantee good general performance, a requirement which drives models to be larger and larger for diminishing returns. A hybrid model can provide sensible defaults for many cases, vastly reducing the need for large statistical models.
“Augmented Engineering will change the way we approach mainstream engineering. It will take people and processes, products and services to new levels of intelligence and efficacy.”
“Hybrid AI brings together generative machine learning models with sources of structured knowledge and symbolic reasoning. It is tailored to the world of engineering, where the data landscape is inherently multi-modal.”
The Hybrid AI approach also helps to address one of the most insidious issues in AI deployments – that AI solutions (particularly Gen AI) fail in different ways to humans, which makes them unpredictable and untrustworthy. The harsh reality is that systems operating in the real world are judged not by how often and how well they succeed, but by how often and how badly they fail. These AI-specific failure modes must be managed, and engineering leaders need to decide on the scientific or engineering models or techniques they can use to control them and boost value.
To a large extent, choosing where to invest and determining what will be acceptable will depend on balancing a number of factors including through-life cost/benefit, competitive separation, what the supply chain is doing, what regulations prescribe, and what markets and society expect. For example, a one-off mechanical design solution that is highly carbon-expensive because of its computation, analysis and data costs may well be a cost worth paying if it reduces the weight of a commercial airliner fleet by 1%.
What ought to emerge from all these considerations is a combined model that is fit for engineering purposes – a model that draws not just on Gen AI, but on other disciplines, other types of AI, and sources of structured knowledge, to create a problem-solving instrument appropriate to the task at hand, that is transparent, trustworthy and reliable, and that delivers truly Augmented Engineering.
