The Impact of Artificial Intelligence on Aluminum Engineering in Mexico

I. From the Foundry to the Algorithm: Knowledge as an Industrial Force

In 1900, when the first furnaces of the Fundidora de Fierro y Acero de Monterrey were lit, Mexico was not merely opening a plant. It was declaring an ambition: to prove it could turn raw material into modern industry. Those furnaces represented physical capital and investment, but also something less visible — accumulated technical knowledge, operational discipline, and the human capacity to make complex systems work 1.

More than a century later, Mexico's aluminum industry continues that tradition. Melting furnaces, extrusion lines, heat treatment, rolling, secondary casting. Solid infrastructure. Experienced engineers. Operators with years — sometimes decades — of practical knowledge.

And yet there is a quiet paradox.

Mexico does not have a fundamental problem of technical capability in manufacturing. It has a problem of formalizing and structurally accumulating technical knowledge. Much of the operational judgment — the reasons behind decisions, the fine adjustments, the deliberate sacrifices of efficiency in favor of quality — remains scattered across individual memories, emails, non-standardized logbooks and hallway conversations.

In an industry where qualifying a product can be the difference between selling to a local market and becoming a supplier to demanding global chains, the advantage does not lie only in the furnace or the press. It lies in the ability to decide correctly under real conditions.

Today, artificial intelligence opens a different possibility: not capturing every data point in the process, but capturing the technical reasoning behind critical decisions.

II. Beyond CAPEX: The Error of Confusing Digitization With Hardware

For years, digital transformation in manufacturing was presented as a matter of heavy infrastructure: additional sensors, integrated systems, complex instrumentation, software tied to hardware as a justification for capital investment. The model was clear: raise CAPEX to enable "Industry 4.0" 2.

In many cases this produced important advances. But it also created cultural resistance, underused systems, and dependence on external technology vendors.

In most Mexican aluminum plants, however, the problem is not a lack of sensors. The critical parameters are already measured. Temperatures are logged. Cycle times are monitored. Scrap is quantified.

What is not captured systematically is the technical reasoning.

When an engineer decides to lower the operating temperature to avoid residual stresses in a particular alloy — even at the cost of thermal efficiency — that decision rarely gets formalized as a repeatable criterion. When extrusion speed is adjusted to improve dimensional stability, that experience seldom becomes structured heritage.

The real opportunity is not in measuring more. It is in thinking better, and documenting that thinking.

III. Cognitive Capacity as an Industrial Asset

In operations science, particularly in frameworks such as Factory Physics 3, installed capacity is usually analyzed through three fundamental variables: inventory, time and physical capacity. Decisions about WIP, lead time and utilization determine financial and operational performance.

But there is a fourth variable that is rarely formalized: cognitive capacity.

Traditionally, talking about capacity means talking about furnaces, presses, shifts, square meters. Yet a plant with engineers able to diagnose deviations quickly has greater real capacity than another with the same physical equipment but without that judgment.

Cognitive capacity is the collective ability to form technical hypotheses under uncertainty, design structured experiments, interpret results with physical judgment, make decisions balancing efficiency, quality and risk, and document and transfer that knowledge.

When a team operates only on tacit experience, that capacity is fragile. It depends on individuals. It is not transferable. It does not scale. When that capacity is formalized through intelligent assistants that help structure experimental design, generate technical memos and record hypotheses alongside results, it becomes an organizational asset.

It is not a replacement for the engineer. It is armor.

IV. Cognitive Augmentation: The Engineer Does Not Experiment Alone

Design of experiments is not new on the plant floor. Many engineers beside furnaces or extrusion lines already run constant iterations, adjusting variables, observing behavior, learning from the physics of the process.

The difference today is that the process can be accompanied by artificial intelligence agents that help structure the DOE, suggest relevant variables from prior history, record explicit hypotheses, generate automatic technical memos, and connect decisions to measurable results.

This is not about capturing thousands of points per second at the lowest operational layer of the ISA-95 architecture 4. It is about capturing the moment of decision: what was observed? what risk was identified? what trade-off was accepted? what result was obtained?

That record turns experience into structured memory. And that memory, accumulated over years, constitutes replicable technical heritage.

V. From Data to Judgment: Cumulative Technical Heritage

Industrial data is voluminous. Technical judgment is scarce.

Flow management on the plant floor is a concrete example. In practice, each workstation operates with a buffer of work in process — an intermediate inventory that, following the principles of Factory Physics 3, is not waste but a regulator of variability. Managing those buffers optimally requires continuous visibility of work in process at every point of the flow: if material accumulates at a bottleneck, the cycle time of every order passing through it grows. Artificial intelligence makes it possible to detect those accumulations in real time, anticipate saturation and regulate material release into the flow, turning what is traditionally managed by the supervisor's intuition into a quantitative control mechanism.

The same logic applies to product qualification. In aluminum, qualifying a product can mean fine adjustments in composition, delicate thermal control, validation of mechanical behavior and compliance with strict specifications. If the knowledge generated in each qualification stays in individual experience, it is lost to staff turnover. If it is structured as a library of technical decisions, the next qualification is faster and less risky.

In both cases — flow regulation and product qualification — the plant stops being merely physical infrastructure and becomes a system of cumulative learning. That changes the nature of the asset.

In a simulation study run with Patok's graph model 7, we evaluated the effect of automatically controlling two variables in a model aluminum plant: when each order is released into the production flow, and the maximum inventory level allowed before each workstation. Using particle swarm optimization (PSO), thousands of possible production plans were evaluated under realistic conditions — including setup times between products, random breakdowns and process variability. The results showed significant improvements in both work-in-process reduction and net profit (Figure 1), confirming that the most powerful lever in flow management is not machine speed but discipline in releasing material.

Figure 1. Cumulative net profit (USD) over a simulated 90-day quarter for three planning strategies: reactive (FIFO, red), deterministic (EDD, yellow) and AI with PSO (green). The PSO strategy achieves the highest profit by the end of the period by simultaneously reducing the financial cost of work in process and lateness penalties. Source: interactive simulator 8.

The decision space explored by the PSO algorithm — more than 10,000 distinct production plans — can be visualized through parallel coordinates (Figure 2), where each line represents a complete plan with its buffer parameters and results. The concentration of green solutions (high profit) in specific ranges of the decision axes confirms that there are optimal regions of the parameter space that manual tuning does not make evident.

Figure 2. Parallel coordinates of the production plans evaluated by PSO. Each line is a distinct plan; color indicates the resulting profit (red = low, green = high). The decision axes include the buffers per machine and the start delay; the result axes include average WIP, on-time delivery and net profit. The highest-profit solutions (green) converge in specific parameter ranges, revealing the structure of the optimal decision space. Source: interactive simulator 8.

The complete interactive study is available as a dedicated white paper 8.

VI. Computer Vision and Passive Capture: Software Over Generic Hardware

In parallel, deploying low-cost industrial cameras makes it possible to capture operational states without massive investment. This is not about exotic sensors, but about software able to visually classify machine states, operational congestion, stoppages never formally logged, and visible variability in flow.

This kind of lightweight instrumentation does not seek to replace the engineer. It seeks to increase operational visibility without cultural friction or complex infrastructure requirements 5.

The value is not in the camera. It is in the operational model that interprets what it observes. Software becomes the layer translating physical reality into actionable information. And when that information connects to a structured memory of decisions, the organization gains coherence between what it knows and what it sees.

VII. Financial Implications: Reducing Structural Risk

A company whose operation depends heavily on key individuals carries greater perceived risk. A company that formalizes its technical judgment reduces that dependency and accelerates knowledge transfer.

This directly affects the start-up time of new lines, the speed of qualifying new products, reliability in the eyes of international customers, and the capacity for geographic expansion.

In the particular case of aluminum, these decisions carry an additional financial dimension: raw material cost is indexed to the London Metal Exchange (LME), a commodities market with daily volatility. The three classic levers of operations science — increase inventory, invest in capacity, or negotiate delivery times — are ultimately financial trade-offs. Every kilogram of aluminum immobilized as work in process carries a cost that fluctuates with the market. With artificial intelligence it becomes possible to quantify those trade-offs in real time: what it costs to hold a buffer at today's price, versus the risk of starving a critical workstation.

CAPEX remains necessary. But the cognitive asset reduces the risk attached to CAPEX. A plant in Brazil, Colombia or the United States can have similar furnaces. What differentiates is the ability to stabilize processes quickly and document that stability. That is intangible capital. And well-structured intangible capital has value.

VIII. Mexico and the Industrial Knowledge Economy

For decades, Mexico competed on cost, geographic location and installed capacity. But the global economy is shifting toward models where structured knowledge carries more weight than infrastructure alone 6.

The question is not whether Mexico can produce aluminum. It already does.

The question is whether it can turn its industrial experience into transferable heritage. If technical decisions, operational criteria and product qualification methodologies are formalized and structured, they can become exportable models. Not confidential data. Not trade secrets. But decision architecture.

That positions Mexican industry not only as an executor of manufacturing, but as a generator of replicable industrial knowledge.

IX. The Factory as a System That Learns

When technical memory is structured, the factory stops operating solely as a physical system. It begins to operate as a cognitive system.

Every documented decision feeds the next. Every structured experiment reduces future uncertainty. Every recorded adjustment prevents repeating a mistake. Installed capacity is no longer only thermal or mechanical. It is cognitive. And when cognitive capacity grows, the organization's total capacity grows with it.

In ongoing research, we have developed through the Patok software 7 a graph-based manufacturing model in which every transformation point — a machine, a warehouse or a shipping point — is represented as a digital node with a defined flow structure: input queue, processing, output inventory and intermediate buffer. Work in process is modeled as work instructions whose content is mass in transformation, and the set of connected nodes forms a directed graph capturing the real topology of the plant. This model lets artificial intelligence agents operate with structured context of the manufacturing flow, making it easier to regulate buffers, detect accumulations and anticipate bottlenecks — functions that traditionally depend on the supervisor's tacit experience.

This is not a technological promise. It is a logical consequence of formalizing learning.

X. Conclusion: From the Furnace to the Heritage

The Fundidora symbolized Mexico's entry into industrial modernity. Today the challenge is different.

Having furnaces and presses is not enough. Measuring more variables is not enough. Acquiring systems is not enough.

The real strategic leap consists of capturing and structuring the technical reasoning that has let Mexican industry compete for decades.

Artificial intelligence does not replace the engineer. It equips them.

And when technical judgment stops being scattered experience and becomes structured memory, the plant does not only produce aluminum. It produces cumulative knowledge.

At that moment, Mexican industry stops competing solely on cost or physical capacity and begins competing on industrial cognitive capacity. And that difference can redefine its place in global manufacturing.

References

1 Cerutti, M. (2000). Propietarios, empresarios y empresa en el norte de México: Monterrey de 1848 a la globalización. Siglo XXI Editores.

2 Schwab, K. (2016). The Fourth Industrial Revolution. World Economic Forum.

3 Hopp, W. J. and Spearman, M. L. (2011). Factory Physics. 3rd edition. Waveland Press.

4 ISA-95 / IEC 62264. Enterprise-Control System Integration. International Society of Automation.

5 McKinsey & Company. (2024). AI in manufacturing: Unlocking the value of data and analytics. McKinsey Global Institute.

6 Deloitte. (2023). Smart factory for smart manufacturing. Deloitte Insights.

7 Canales Siller, H. (2025). Patok: operational visibility and artificial intelligence platform for manufacturing. Monterrey, Mexico. https://patok.in

8 Canales Siller, H. (2025). Manufacturing flow optimization through WIP release control. Patok Research WP-001. https://patok.in/en/research/manufacturing-flow-optimization

Dr. Horacio Canales Siller is an engineer with a doctorate in materials science, specialized in aluminum and its alloys. He is the founder of Patok, an operational visibility and artificial intelligence platform for industrial plants in Mexico.