AI for engineering: Why control and systems thinking matter

We are sleepwalking into a skills gap in Artificial Intelligence (AI). A growing number of AI practitioners are trained almost exclusively by computer scientists, adept at deploying open-source pre-coded libraries and producing models that perform impressively on artificial demos and benchmarks. On paper, these practitioners look great. However, in practice, many lack a deep understanding of what is happening under the hood: the optimisation and control theory principles that provide the necessary insights for real-world deployment. As such, industry is increasingly looking past candidates with generic AI skills in favour of those with a deeper understanding.

To dig deeper into this emerging skills gap, we must first agree on what we mean by AI. It is not merely generative text or images. AI is a class of methods capable of solving tasks at or beyond human capabilities, without requiring real-time human input or hand-crafted decision rules. Within this definition, we can identify two distinct philosophies. The first is inductive, coming from a dataset (Machine Learning). The second is deductive, coming from physical models (Control Theory).

Advertisement

Historically, Control Theory and Machine Learning have approached the same class of problems from opposite directions. Control theory, which dates back at least to James Clerk Maxwell’s 1868 analysis of the steam engine governor, predates digital computation and the era of large datasets. Lacking the luxuries of abundant data and cheap computation, it has nonetheless still been successful at solving precisely the same problems of safety and reliability that confront today’s AI systems. Engineers used physical laws (like Newton’s laws of motion) to derive models and design controllers (policies) with guarantees that the system will behave acceptably.

The downside of the control theoretic approach has been that deriving accurate models for complex, interconnected systems is often extremely difficult, if not impossible. While there exist methods to partially mitigate modelling error, the resulting optimisation problems are frequently high-dimensional and computationally intractable. Without modularisation, these approaches scale poorly when applied to systems of national scale, such as power grids and transport networks. Systems with poorly understood models, such as language and human behaviour, present even greater challenges. Moreover, control theoretic solutions tend to be slow to iterate and deploy, requiring careful analysis, tuning, and certification. Control theory draws on an unusually broad and demanding mathematical toolkit, creating a steep barrier to entry that limits the pool of practitioners able to apply it effectively.

In contrast, machine learning has proven highly effective in applications that allow for the collection of large data sets, but resist low-complexity model descriptions, where control theory is most effective. The primary advantage here is that we do not need to understand how the system works to automate it. This creates a low barrier for entry where practitioners can immediately start to create models based on open-source data sets and toolboxes. With vast enough data, such as how LLMs are trained on next token prediction given archived text from the entire internet, models begin to exhibit emergent capability, like writing code, that surprises even their creators.

However, this data-driven strength is also its Achilles’ heel. Some applications do not have and will never have large enough data sets that are required to learn useful prediction models, think of the black swan events in financial markets. Machine learning lacks the data sample efficiency of humans; whereas a human can learn to drive safely in 20 hours, pure reinforcement learning agents often require millions of simulated miles to achieve basic competence. Moreover, the machine learning models are commonly static; they capture the statistical correlations of the past, not the causal dynamics of the future. A model trained on data from 2025 may fail catastrophically if the physical environment shifts in 2026, as the weights do not automatically update to reflect new realities. While a hallucination in a chatbot is a manageable nuisance, a similar error in a smart grid or an autonomous vehicle is a physical hazard. Consequently, without the safety guarantees provided by control theory, purely data-driven AI remains difficult to trust in high-stakes tasks where lives are at risk.

This dichotomy presents an engineering grand challenge: how do we fuse the two distinct fields of machine learning and control theory together to enjoy the advantages of both. The challenge ahead is as much educational as it is technological. The future belongs to engineers who are bilingual in both disciplines. Students should seek university programmes that merge both fields, providing a comprehensive and well-rounded skillset. For India, this message should resonate strongly. As the nation continues to push towards modernising its large manufacturing sector, the need for automation grows stronger. Demand for a skillset in ‘AI for Engineering’ is set to skyrocket.

(The writer is a Lecturer in Machine Learning and Control Theory and Programme Lead for the MSc Artificial Intelligence for Engineering, The University of Sheffield, UK)