When Apples Fall and Models Fail
Mohamed Gabry, Ph.D.
Petroleum Engineer|Reservoir Management|Geo-mechanics|Well Completions|Well Intervention|Production Operations|Hydraulic Fracture|Acid Stimulation|Data Analysis|Machine Learning|Wavelet Transform
Engineering, Assumptions, and AI
Throughout human history, science has undergone cycles of advancement and regression—eras marked by humanity's relentless pursuit to understand the world around us. These cycles reflect our repeated efforts to decode nature’s complexity. Remarkably, ancient civilizations like the Egyptians exhibited technological sophistication that still puzzles modern scholars, suggesting that progress in human knowledge is not always linear.
The universe operates according to fundamental laws of physics and chemistry—whether classical or quantum—regardless of human awareness or interpretation. An apple will fall to the ground whether we describe it using Newton’s law of gravity or Einstein’s theory of relativity. To make sense of such intricacies, humans often rely on fixed models and assumptions. The "KISS" principle—Keep It Simple, Stupid!—captures our tendency to simplify because our cognitive capacity struggles with highly multivariable systems. As variables and their interdependencies increase, the resulting mathematical models quickly grow more complex and less intuitive.
A key distinction between science and engineering lies in their goals: science aims to uncover and understand principles with precision, while engineering seeks to apply those principles to solve real-world problems. Scientists often resist assumptions to preserve nuance, whereas engineers embrace them as tools for building practical solutions within specific constraints. Every engineering formula is valid only within the boundaries defined by its assumptions. Unfortunately, these foundational assumptions are often overlooked in modern engineering education. Many engineers apply formulas routinely, forgetting the conditions under which they were derived. This neglect usually remains invisible—until an edge case breaks the model.
As we gained deeper insights into how the human brain works—particularly how infants acquire language, logic, and emotional intelligence—we began designing machine learning models to mimic this learning process. Today, such models help us navigate vast datasets with unprecedented speed and efficiency. But they’re not without limitations. The data we feed into them is often noisy, biased by physical assumptions, riddled with outliers, and subject to human error. Still, by extending our cognitive reach through machine learning, we are refining our ability to model, interpret, and predict the complexities of the natural world.
However, even machine learning models are built on assumptions—especially when we arbitrarily designate certain features as inputs and others as outputs. If these designations are rooted in simplified interpretations, the resulting models can inherit those biases. Consider a model trained to predict an output feature Y from input features x1, x2, x3, and x4. If x1 and x2 are interpreted using different methodologies—Z1 and Z2—based on differing assumptions F1 and F2, the model becomes bound by those assumptions. Applying it outside that context can produce unreliable or misleading predictions. This highlights one of the major strengths of pure machine learning approaches over hybrid physics-based models: they can learn patterns without being constrained by predefined physical assumptions. It also explains why real field data, which captures the full natural complexity, is often more valuable than synthetic data generated from simulators—data that inherently reflects the assumptions and constraints of the simulation environment.
This brings us to a critical point: before building machine learning models, we must first conduct a rigorous system analysis—especially in fields like subsurface engineering. Applying 'Signal and System' techniques allows us to understand the behavior of physical systems holistically. Unfortunately, many current machine learning efforts skip this foundational step, leading to models that predict one parameter of a physical system from another—or worse, from completely irrelevant features. A model might, for instance, correlate the success of a hydraulic fracturing job with the name of the frac engineer’s neighbor. Statistically, the model could appear accurate—but scientifically, it's nonsensical. Without grounding models in system-level understanding, machine learning can produce spurious correlations that mislead rather than inform.
In a world increasingly driven by AI, this integration of domain knowledge and data science is no longer optional—it’s essential.
That’s precisely why, at the University of Houston in research group of Prof. Dr. Mohamed Y. Soliman, PhD, PE, NAI , we’re focused on developing generic, physics-respecting solutions tailored for machine learning models—solutions using Wavelet Transform that are only possible after applying rigorous 'Signal and System' techniques to analyze the complex behavior of subsurface engineering systems. Rather than rushing into data-driven modeling, we begin by understanding the system’s dynamics, constraints, and signal behavior.
Kudos to the Current Team Members:
and Previous Team Members:-
Recommended by LinkedIn
Why does this matter? Because without this foundation, even the most accurate-looking ML models can become misleading or brittle when applied in real-world scenarios. So here's the question: Shouldn’t all machine learning in engineering start with understanding the system first? Let’s talk about it.
📄 Read more about our work in these papers:
Section Head @ Khalda Petroleum Company (Apache) | Petroleum Engineer
1moإنني معجب بهذا، Mohamed
Senior Petroleum Engineer, Rig-less Operations and Artificial Lift Systems; Hydraulic Jet Pumping, Sucker Rod Pumping, ESP and PCP │ IWCF Certified │ NEBOSH IGC Certified.
1moI really enjoyed the depth and clarity with which you connected engineering fundamentals, system thinking, and modern AI challenges. Thanks for sharing such a thoughtful perspective
Experienced in AI application for Oil Industry | Master's Student - Digital Oilfield Technologies - University of Southern California
1moFully agree