UM6P CMSI — AI & National Security Seminar Intelligence 2026
Purpose: Intelligence extracted from a live seminar at UM6P Ben Guerir campus, April 25, 2026 (Math & Maroc YouTube channel, @mathmaroc1396). Two confirmed speakers: Ahmed Ratnani (UM6P Vanguard Center director, digital twins + OT) and Hamza Keurti (UM6P College of Computing, geometric deep learning). Both are inside the OCP Group institutional ecosystem. Source: Full verbatim transcript, timestamped (22:13 → 2:28:48), captured 2026-04-26 Related: OCP Group · IEIA · IROC · MaestroHub — Intelligence & Strategic Assessment 2026 · Africa Industrial IIoT Market Intelligence 2026 · Morocco Industry 4.0
⚡ Why This Matters for IEIA
This is not a generic academic seminar. Two speakers, both at UM6P, embedded in the OCP Group institutional ecosystem. Both are research-level contacts with elite international backgrounds who are now building Morocco’s AI/OT academic infrastructure — the intellectual layer above the industry IEIA targets.
Speaker 1 — Ahmed Ratnani — OT-focused, UM6P Vanguard Center director, explicitly named OCP digital twins:
- He explicitly named OCP Group digital twins as a strategic priority — “if we look at digital twins for companies like OCP, if we are able to link the digital twins that are produced for production to organizational digital twins…”
- He moved from IT to OT — “I moved to the OT side, operational technology side rather than IT, for many reasons because for industry this is what we need and also for defense this is what we need.” This is IEIA’s exact positioning.
- He directs the UM6P Vanguard Center (Digital Twins + Cyber-Physical Systems Division) — a new UM6P center he described as “a very young college.” He said: “feel free to reach us if you want to know more about our vision.”
- He was interim director of CYSA (Cyber Security for Africa, also at UM6P).
- CERN–UM6P Partnership team lead — gives him international credibility and access to European research infrastructure that IEIA could leverage.
- PhD from Strasbourg University, prepared at INRIA — French-language academic background = cultural fit with IEIA’s Moroccan/French approach.
- Former researcher at Max-Planck Institut für Plasmaphysik, Munich — high-caliber international career now anchored at OCP’s own university.
- Email confirmed: ahmed.ratnani@um6p.ma — ready to contact now.
IEIA action: Ahmed Ratnani directs UM6P’s digital twins / cyber-physical systems unit — which is literally building the academic infrastructure layer for OCP Group’s digital transformation. He is not a purchasing decision-maker but is likely the strongest academic bridge to OCP’s digital leadership (CDO, VP Digital) in all of Morocco.
Speaker 1 Profile — OT / Digital Twins / Vanguard Center ✅ CONFIRMED
✅ Identity confirmed: Ahmed Ratnani — confirmed via web search (ratnania.github.io, LinkedIn, Google Scholar, ResearchGate, GitHub). “Professor Ahmed” = given name; “Mu Ahmed” = colloquial “Moulay Ahmed.”
| Field | Detail |
|---|---|
| Name | Ahmed Ratnani ✅ |
| Title | Assistant Professor, Applied Mathematics & Computer Science — UM6P |
| ahmed.ratnani@um6p.ma | |
| Website | ratnania.github.io |
| https://www.linkedin.com/in/ahmed-ratnani-0a77344/ | |
| GitHub | github.com/ratnania |
| Google Scholar | 1,006 citations · H-index 16 · i10-index 24 · Interests: Model Reduction, Computational Physics, HPC, Applied Linear Algebra, Finite Elements |
| Current unit | Director, UM6P Vanguard Center — Digital Twins and Cyber-Physical Systems Division (the “CMSI” reference in transcript = this center, described as “a very young college”) |
| Other UM6P roles | Director (interim), CYSA (Cyber Security for Africa — “CISA” was auto-caption misread) · Director, Al-Khwarizmi Department · Team Leader, CERN–UM6P Partnership · Holder, Multiphysics HPC Chair |
| Current focus | OT side (Operational Technology) — explicitly stated transition from IT/cyber to OT; Vanguard Center’s Digital Twins + Cyber-Physical Systems Division |
| PhD | Applied Mathematics / IsoGeometric Analysis in Plasma Physics & Electromagnetism — University of Strasbourg, prepared at INRIA (2008–2011) ✓ |
| Post-doc | Researcher, Max-Planck Institut für Plasmaphysik, Munich — led the Magnetohydrodynamics research group (2014–2019) ✓ |
| Research areas | IsoGeometric Analysis, Model Order Reduction, HPC, numerical methods for complex systems, digital twins, cyber-physical systems |
| Contact invitation | ”Feel free to reach us if you want to know more about our vision” — explicit open door |
| Vision | ”The Modern Mathematical Moroccan Mathematical School” — Morocco-relevant research, not imported academic agendas |
| Academia-industry thesis | Digital twins must be co-developed between academia and industry. UM6P Vanguard Center = “the in between.” |
Speaker 2 Profile — ML / Geometric Deep Learning / Cognitive Science ✅ CONFIRMED
✅ Identity confirmed: Hamza Keurti — confirmed via Gemini video metadata + web search (hamzakeurti.github.io, LinkedIn, ResearchGate, GitHub).
| Field | Detail |
|---|---|
| Name | Hamza Keurti ✅ |
| Title | Assistant Professor, Machine Learning — UM6P College of Computing |
| hamza.keurti@um6p.ma | |
| Website | hamzakeurti.github.io |
| Confirmed (Hamza Keurti — UM6P) | |
| GitHub | hamzakeurti |
| PhD | ETH Zürich + Max Planck Institute for Intelligent Systems — as a fellow of the Max Planck ETH Center for Learning Systems |
| PhD Advisors | Benjamin Grewe (ETH Zürich, Neural Learning and Intelligent Systems Lab) + Bernhard Schölkopf (Max Planck Institute for Intelligent Systems, Director — Empirical Inference) |
| Engineering degree | Diplôme d’Ingénieur — Centrale Paris (École Centrale Paris) + MSc Advanced Computing — Tsinghua University (China; auto-captions rendered as “Chinua University”) |
| Industry experience | Internship at SLB (Schlumberger) |
| Research focus | Intersection of machine learning, algebra, and cognitive science; structured and compositional representation learning; symmetry discovery through agent-environment interaction; world models with geometric/equivariant constraints |
| Key publication | Homomorphism AutoEncoder (ICML 2023) — learning group-structured latent representations via action-observation pairs |
| Talk title | ”Algebra at the Core of Artificial Intelligence: from Geometric Deep Learning to Representation Learning” |
| Digital twins connection | Opened talk with: “I want you to keep in mind this idea of digital twins. I’ll tell you why in a bit.” Defined the brain as “a mental twin of the outside world.” His world models = AI perceiving factories the way a human operator does. |
Why Hamza Keurti matters for IEIA:
- ETH Zürich + Max Planck (Schölkopf lab) = world’s top-tier AI research pedigree — Schölkopf is one of the 5 most cited ML researchers globally
- Geometric deep learning / equivariant representation learning = the mathematical foundation of the AI reasoning layer that sits above the UNS data layer IEIA builds
- His ICML 2023 work (Homomorphism AutoEncoder) is exactly the architecture that could perceive OCP plant states from UNS data streams
- Explicit digital twin framing in his talk = same conceptual vocabulary as IEIA’s OCP pitch
- Email confirmed → direct outreach possible immediately
- At UM6P College of Computing = institutional bridge to CMSI, CISA, and OCP’s digital infrastructure
Organizing Association — Math & Maroc
Name confirmed as “Math & Maroc” — YouTube channel @mathmaroc1396, confirmed by Gemini. NOT “MATEMAROC.” The association has no MATEMAROC handle confirmed anywhere.
| Field | Detail |
|---|---|
| Name | Math & Maroc (moderator transcript phonetic: “Mate Maruk / Matarak”) |
| YouTube | @mathmaroc1396 |
| Founded | 2016 — to help Olympiad candidates compete internationally |
| Event created | MATM Competition (annual, July 24–28, College of Computing) — for postbac/undergraduate students |
| Members | ~200 (as of April 2026), growing |
| Events | 0 in 2022 → 1 in 2023 → ~10 in 2024–26 |
| Divisions | Seminar division (organized this event), competition division |
| Active (announced competitions) | |
| Volunteers | Always recruiting — “if you are interested please apply” |
| Base | College of Computing, UM6P campus, Ben Guerir |
| Video published | April 25, 2026 — the seminar was uploaded to the Math & Maroc YouTube channel the day after this transcript was captured |
IEIA relevance: Math & Maroc is a student mathematics association embedded in UM6P’s ecosystem. 200 members = future OCP engineers, researchers, and digital transformation practitioners. Sponsoring or speaking at a Math & Maroc seminar gives IEIA access to the next generation of OCP’s technical workforce.
Full Transcript — Key Extracts & Intelligence
AI and National Security — Core Thesis
“National security is not about defense. It is all the sectors that are important for the country. Education is important. Energy — we are living it today with what’s happening in the Middle East. Agriculture for sure and water. Economics and industrial systems as well. Cybersecurity for sure, health and biosecurity. Military and defense are at its core but there are many many other fields that are relevant to national security.”
IEIA framing: This is the exact framing that makes MaestroHub / UNS infrastructure a national security argument, not just an industrial efficiency argument. OCP’s phosphate is Morocco’s strategic resource. Securing OCP’s digital infrastructure = national security. IEIA can use this framing in government-adjacent conversations.
OCP Digital Twins — Direct Mention
“If we look at digital twins for companies like OCP, if we are able to link the digital twins that are produced for production to organizational digital twins, then we are able to completely simulate the structuring of the company and see what is the effect on the market and how the feedback loop is done.”
Intelligence: The speaker is describing a production digital twin (shop floor, SCADA, PLCs) connected to an organizational digital twin (ERP, planning, market modeling). This is exactly the IROC architecture: real-time plant data (UNS layer = MaestroHub) feeding into an organizational intelligence layer (SAP S/4HANA + AI). IEIA builds the production digital twin. The organizational twin comes from SAP/AI layers above it.
This is the first independent public confirmation from inside the UM6P ecosystem that OCP-scale digital twins are a stated strategic priority at the institutional/academic level — not just an OCP internal initiative.
IT → OT Transition (Speaker’s Own Career)
“I moved to the OT side, the operational technology side rather than IT, for many reasons because for industry this is what we need and also for defense this is what we need.”
Intelligence: Even the academic leadership at UM6P’s new CMSI is pivoting toward OT. The direction of Morocco’s applied AI research is toward operational systems — not purely theoretical or IT-side. IEIA is already on the right side of this pivot.
The Talent Argument (Moroccan Math as Strategic Asset)
“Morocco is really entering a new era where for the first time in its history we can make use of something that we are very good at which is mathematics.” “Most of the students who are in France like Polytechnique, Centrale — now they’re aiming to do their PhD, so it is a natural continuation today.” “Moroccan talents can help Morocco to be positioned in a very interesting role across the world.”
IEIA framing: IEIA employs Moroccan engineering talent in OT/IIoT. This framing gives IEIA a national narrative: Moroccan mathematical talent + OT infrastructure knowledge + Arabic/French cultural context = uniquely Moroccan industrial digital sovereignty capability. No foreign SI can say this.
Digital Twins — Market Numbers (Quotable)
“The global market for digital twins is already 13 billion by 2030.” “Agriculture: yield increases 10–25%, water optimization reduction up to 40%.”
These are validated academic-level statistics. IEIA can cite them in proposals with attribution to “UM6P research community” or the speaker specifically.
Education as National Security — The Junior Engineer Problem
“Today a lot of companies are avoiding hiring junior engineers because with code and other tools they can do what a junior engineer does. And this is very problematic because how can you find your way in the job market if we remove the first step which is being a junior.”
Intelligence: This is the talent pipeline problem. IEIA can position itself as a company that still hires and develops junior OT engineers — a national talent development argument on top of the commercial one.
The Value-at-Risk Warning (AI Safety Framing)
“You have a precision of 99%. And for this 1% of error you are not computing the risk. Nobody computes the value at risk that you are exposed to when your algorithms fail.” “We are in AI doing exactly what caused the financial crisis of 2007. People didn’t compute the value at risk.” “Ariane 5 — the rocket self-destructed 40 seconds after launch.” “Patriot failure in ‘91 — 28 soldiers killed because of an error of 10⁻⁷ in computations.”
IEIA use: This framing supports the argument for governed, auditable UNS infrastructure over ad-hoc AI/IoT deployments. The speaker is making the case that unvalidated AI in critical systems is dangerous — which is exactly why you need a governed data layer (MaestroHub) before deploying AI on top. IEIA can use this in OCP conversations: “Before you run AI on your plant data, you need data you can trust. That’s the UNS layer.”
Cyber Physical Attack — The Lebanon/Beeper Reference
“I think you remember what happened in Lebanon with the attack on Hezbollah through beepers. This was a cyber physical attack. We can control your device from very far away because in every electronic device there are what we call back doors.” “Morocco does not produce any electronic device and one of these days we can be hacked because we have a very large attack surface.”
Intelligence: The speaker is making a strategic vulnerability argument about Morocco’s dependence on foreign hardware. For IEIA, this argues for open-architecture, software-defined OT infrastructure (MaestroHub, open-source protocols) over proprietary hardware-locked systems (like 4Sight Holdings’ AVEVA stack or Siemens proprietary gateways). Open stack = no single vendor back door.
Defense AI — Numbers
“Global military AI spending: 40 billion in 2028.” “Drone market reached almost 3 million to shoot down a drone that costs $10,000.”
AI as a Military Domain
“AI is not just a tool. AI is a domain. In military language, a domain means like air, like ocean — everywhere where a war can happen. AI now is a domain. It has been defined as a domain, not as a tool.”
Two Timelines of AI (Civil vs Military)
| Year | Civil AI | Military AI |
|---|---|---|
| 1956 | Dartmouth workshop — AI birth | — |
| 1958 | Expert systems begin | ARPA/DARPA created (avoid strategic surprises) |
| 1980s | Expert systems in industry | SGE — first air defense C2 system operational |
| 2012 | AlexNet / ImageNet breakthrough | — |
| 2016 | AlphaGo defeats Lee Sedol | Project Maven launched (AI for ISR intelligence surveillance) |
| 2010–now | ChatGPT, LLMs, generative AI | Maven still active (used in Iran operations per speaker) |
| 2004 | — | Stuxnet — first known cyber physical weapon (Iran nuclear) |
Strategic read: Military AI has been ahead of civil AI at every step. DARPA preceded OpenAI by 50+ years. The speaker is building the case that Morocco needs to develop its own AI/OT capabilities now — before the gap with advanced militaries becomes unbridgeable.
Event Context
| Field | Detail |
|---|---|
| Event type | Academic seminar — AI and National Security |
| Organizer | Math & Maroc (mathematics student association, @mathmaroc1396) + possible IEIA co-organization (“thanks to a collaboration made by Mate Maruk”) |
| Second speaker mentioned | Not yet captured in transcript (seminar had two speakers) |
| Audience | Students, researchers, young professionals — Moroccan university context |
| Language | English (seminar) with French references |
| Location | Likely UM6P campus or partner institution — College of Computing mentioned |
| Date | April 25, 2026 — confirmed (Gemini confirmed video upload date; recorded at UM6P Ben Guerir campus) |
Note on “collaboration made by Mate Maruk”: The moderator said “thanks to a collaboration made by Mate Maruk that I will leave Mr. Muhammad to tell you more about.” This may mean IEIA co-organized or co-sponsored this seminar with Math & Maroc. If so, IEIA already has a relationship with this event series — confirm with IEIA team internally.
Q&A Session — Additional Intelligence
Digital Twins Are Living Systems (Not One-Shot Projects)
Question: “How do we make sure the digital twins we build today based on yesterday’s data can predict future complexities?”
“Digital twins are not one-shot projects. They are living things. Once you start deploying your digital twin, this means that you should be working on it for the rest of your life. It’s like if you build an industrial plant — you cannot just put it down and leave the site and say it will work on its own. You should improve it. You should maintain it. You should revise all the methods again and again.”
“What is great is that the maths are evolving, computer science is evolving — this will help us improve and upgrade digital twins. But only if we think of digital twins as real ecosystems that should be shared and co-developed between academia and industry. If we put them at the industry level, they will be outdated very soon. If we put them at the academia, nothing will happen at all. They will not even be built. We need something in between. This is what we are trying to do at M6P.”
IEIA use: This is the managed service argument. A digital twin / UNS deployment at OCP cannot be handed off and left. IEIA’s ongoing managed service contract is not optional — it is structurally required. The professor from inside OCP’s university just made this argument for IEIA. Quote him.
Morocco as “Copycat” — Research Sovereignty Critique
“We do not create — not only in law, in everything, almost everything. We are just copypasters. Copycats basically.”
“Take all the PhDs defended in Morocco. Link PhD students to their advisors. Look at what subjects are studied. You will see all the subjects were exported from outside — because a professor did his PhD in Belgium, France, Canada, and came back still working with that foreign laboratory. We are not working on the subjects that matter for us.”
“When NLP started before LLMs, people thought ‘we should have our own systems for Darija’ and everybody started working on Darija. Today nobody is using that work. We should ask ourselves the right questions and work on research subjects that are meaningful for us.”
“We are building the skills that other countries need. If we keep on like this, we will never be an emergent country. We will always be in service for a given country — and we’ll just change this country for another.”
IEIA use: IEIA is building Morocco-specific industrial capability using Moroccan engineers working on Moroccan industrial problems (OCP phosphate, CBAM, water, energy). This directly answers the professor’s critique — IEIA is not copying a foreign business model, it is solving Morocco’s real industrial problem. Use this framing in government-level conversations and in OCP narratives.
AI Bubble Question
“Are we in a bubble? The market says: in one week, stock options reduced by 1,000 billion dollars for major technology companies. So… from our perspective, we are not even small players — we do not exist at all. You can have many publications in the field, you still do not exist. What we should do is think about this because it is a national security problem. AI is here, it will not leave. We should live with it in every sector.”
“Today you should make bets and take risks. If you forget about it today and say ‘I do things as usual, nothing will happen’ — in very few years you will find yourself outdated and nobody will come to you anymore.”
IEIA use: The timing argument for OCP. OCP is not “in the bubble” — OCP’s phosphate is a structural global necessity. The AI/data layer on top of OCP’s operations is not a speculative tech bet — it is operational resilience. The professor’s framing supports urgency.
AI in Law — The Gap Problem
A tech-turned-law student asked about AI regulation in Morocco:
“Engineers do not understand anything in law. People studying law do not understand anything in AI. The gap is so large because of the choices we made in education. Today we should have only scientific people [as a base], and after high school you choose whether you want to do literature or Islamic studies — because we need a scientific mindset.”
Relevance for IEIA: The CBAM compliance architecture IEIA builds is exactly the intersection of engineering (UNS data pipeline) and regulatory law (EU carbon border adjustment). IEIA occupies the gap this professor describes. No pure lawyer can build the CBAM pipeline. No pure engineer understands EU trade law. IEIA does both.
Speaker 2 — Talk Opening: Perception, Representation, and Digital Twins
“I want you to keep in mind this idea of digital twins. I’ll tell you why in a bit.”
“My biggest curiosity has been our perception system. How do we perceive? How did we learn to perceive? Your brain is completely disconnected from the outside world. And through your eyes, somehow it gets the feeling that it has access to the information — how this scene is structured, the depth of this room. There is a lot of inference going on, and it happens seamlessly.”
“My belief is that we are doing exactly what you’ve been describing — we have a sort of mental twin of the outside world. The complex system we’re modeling is the world outside, and we’re maintaining this simulation of what’s going on — to be able to keep up, to think and plan and do.”
“We want to have AI that perceives the way we do — that is as efficient, as intelligent. The question is: through what principles can we have AI that perceives as well?”
Connection to IEIA / UNS: The mental twin framing = the Unified Namespace thesis stated from first principles. The UNS is a digital twin of the factory’s data reality. MaestroHub’s dependency graph is the ontology that lets an AI agent “perceive” the plant the way a human operator does. Speaker 2 is working on the foundational mathematics of exactly the AI layer that sits above IEIA’s UNS deployment.
Speaker 2 — Full Lecture: Algebra at the Core of AI
Talk title: “Algebra at the Core of Artificial Intelligence: from Geometric Deep Learning to Representation Learning” Core thesis: Abstraction — formalized through group theory and equivariance — is what allows intelligence (human or AI) to generalize from limited data. A good digital twin is not a complete replica but an abstracted model that keeps only what matters for the task.
The Problem: Dimensionality Curse
The fundamental barrier to learning from data: to approximate a black-box function by covering input space at density 0.1:
- 1D input → 10 samples needed
- 2D input → 100 samples
- Image space (256×256 RGB) → astronomically many samples → impossible to cover
“We cannot see enough data. This is basically impossible. So there is something else going on — at least for us, because we’re able to perceive so well and we haven’t seen so much data.”
Banana adversarial example: A large AI model correctly identifies a banana. Rotate it, color it red, change background → still recognizes it. Tilt it slightly → identifies it as a “hook” with banana at 1.1% confidence. The model does not understand which changes matter and which do not.
“The AI model does not understand the concepts of a banana — or what changes matter and what changes do not matter.”
Two resulting questions:
- What allows data efficiency — learning from limited samples?
- What allows generalization — performing well outside the training distribution?
His answer: Abstraction — and the mathematics of abstraction is group theory.
The Digital Twin Modeling Philosophy
“When we want to make a model of some reality — a digital twin for example…”
He cited a famous quote (attributed to Norbert Wiener / Arturo Rosenblueth):
“The best material model of a cat is another, or preferably the same, exact cat.”
The joke: If you want a perfect model, just use the original. But then you haven’t gained anything — you’re back to the initial problem with no simplification.
“If you want to make a digital twin of a very complex system, some people have this ideal of simulating every process that happens in this system. But something breaks — it does not allow you to summarize information, to get access to important information for a given task. If you have the whole complexity in your model, you’re back to the same problem. You don’t have a simplification that allows you to see clear patterns.”
What a good model is:
“A good model drops some of the irrelevant information and only keeps a sort of skeleton — the core skeleton of what’s needed for the task.”
The building analogy: A single building has multiple models depending on the task:
- Mechanical load diagram → keeps only what matters for structural integrity; discards electrical, plumbing
- Electrical wiring diagram → inverse
- Plumbing diagram → inverse
“Each of these models is basically shedding away some of the complexity of the initial system. This is one way of seeing abstraction — modeling.”
IEIA strategic use: This is the philosophical justification for why a UNS is not a data lake, not a full simulation, and not a copy of the plant. A UNS is a task-specific abstraction of the plant’s operational reality. The Semantic Model library (77 models in MaestroHub) is exactly this: each model (Energy Cost Mapping, Machine State OEE, ISA-95 Job Order) is a skeleton that keeps what matters for a specific task and discards the rest. Speaker 2’s mathematics formalizes what IEIA builds.
Group Theory → Group Actions → Orbits → Data Efficiency
Groups (review — known to the audience from high school):
- Set of abstract objects with internal composition
- Identity element, inverses, associativity
- Example: 6-element group = permutations of 3 elements = symmetries of a triangle (3 rotations + 3 reflections)
Homomorphisms: Maps between groups that preserve composition structure.
“This is where we stopped in high school — at homomorphisms — right where it gets interesting. We did not see group actions, quotient spaces, equivalence maps. They don’t even tell us where to start our research.”
Group Actions (the key step beyond high school): Instead of abstract elements, assign each group element a transformation of some space:
- Abstract 6-element group → acts on: a set of 3 dots, a triangle, a 3D space
- All three share the same structure — the same group G acts on each
“I have three different objects that share the same structure. It’s the same G acting on each one of them.”
Orbits:
“Take your input space. Keep track of a point as it transforms under the action of every element in the group. It describes a set — this set is an orbit.”
In R² under 2D rotation: every orbit is a circle centered at the origin.
Quotient Space: Collapse each orbit to a single representative point → the space of equivalence classes.
- R² under 2D rotations → quotient space = R⁺ (just the radius — one half-line)
- You went from 2D to 1D by discarding rotation (which doesn’t change the percept)
“I’ve reduced R² to just half a line. I don’t have to cover the whole input space anymore — I just need to approximate the function in this particular quotient space and I can conclude what happens elsewhere.”
Physics analogy (electrostatics): If charge distribution has rotational symmetry → electric field is radial (invariant under rotation) and direction is equivariant (rotates with position). The full 3D computation collapses to studying intensity along a single line (radius only).
Equivariance — The Key Property for Data-Efficient AI
Definition: A mapping f: X → Y is equivariant under group G if transforming the input and then mapping is the same as mapping and then transforming the output.
f(g · x) = g · f(x) for all g ∈ G, x ∈ X
Visually: the commutative diagram closes — you reach the same output whether you transform-then-map or map-then-transform.
Invariance is a special case where G acts trivially on the output space (the output doesn’t change regardless of the transformation — e.g., the label “banana” doesn’t change with rotation).
Why equivariance gives data efficiency:
“I only need to know f(x) — the image of one input — and I can conclude what will be the output for all of g·x — the entire orbit of x.”
This is why humans can recognize a banana at any angle from a handful of examples. The brain has built-in equivariance priors (translation, rotation, scale). Current deep learning models lack this — which is why the slightly-tilted banana becomes a hook.
The research program: How do we build AI models with the right equivariance structure? This is geometric deep learning — the field Speaker 2 works in.
Key Quotes for IEIA Use
“A good model drops the irrelevant information and keeps only the core skeleton of what’s needed for the task.” — justifies UNS semantic models over data lakes
“If you have the whole complexity in your model, you’re back to the same problem — you don’t have a simplification that allows you to see clear patterns.” — justifies abstraction-first architecture
“The same structure — the same G — acts on each of these different objects.” — justifies why the same UNS ontology (ISA-95) works across different plant types
“We stopped at homomorphisms — right where it gets interesting. We did not see group actions, quotient spaces.” — the education gap he’s addressing at UM6P; IEIA’s students trained in this math are a competitive advantage
Hamza’s Actual Research — Symmetry Discovery Through Interaction
This is the part of the talk where he moved from tutorial content to his own research:
“What I’m looking into is the fact that images and input signals are not the first initial source of information. The world does not start from the images on the retina — it starts outside of you. Your eye is just measuring the state of the world and you’re working with the measurements. You don’t have access to the state of the world. You only have access to the measurements.”
The research problem:
- Standard geometric deep learning: you know the symmetry group (rotations, translations) and build it into the model
- Hamza’s research: you don’t know the symmetry group — you have to discover it through interaction
The method:
“For that, what I’m doing is I’m trying to learn the same way humans do — through interaction. You have actuation at some points, an action put into the environment to make it transform, and then you see the effect of your actions materialize in the state of the world — but you don’t have access to the state, only the measurements. I’m trying to use both the observations and our ability to act — our knowledge of the actions we perform — in order to learn equivariant mental twins of the physical world.”
Two structures he investigates:
- Objectness — the world has a compositional structure; it is made of multiple objects
- Temporal dynamics — the world is under the action of time; this makes dynamics happen
World models connection:
“A world model is just trying to replicate how your agent’s actions impact the world — you simulate, you have a digital twin of the outside world. My approach is a special flavor of a world model, but here the action takes the form of a transformation that transforms the input space — which is not something normally in world models. This is more constraining but when it is true, you have a much better representation.”
Compute efficiency argument:
“When I say I don’t need as much data, it also means I don’t need as much compute. You’ve heard about the amount of energy LLM companies are consuming — now companies are investing in nuclear reactors next to their data centers to power training of their next models. That’s completely brute force. What I’m trying to do is: use the symmetries that we know. The problem is we don’t always know the symmetries.”
Category theory connection (student Q&A):
“This is just a special case of a functor between two categories. A good representation would be one that defines a natural transformation from the input category to the representation category. There is definitely effort happening in that direction.”
On simulating human intelligence (student Q&A):
“If your goal is really to create another human, you’ll have to simulate his whole evolution — cultural and everything — and simulate death and hunger and the need to procreate. I mean, people fail at just the perception part. I don’t think we’ll ever be able to simulate the whole thing. You can do abstractions for specific tasks — if you want to simulate the heart and prevent disease, you can do that. But trying to replicate the whole thing — I don’t see the point, and I don’t think it’s feasible.”
Reference to Human Brain Project:
“This was a huge project in French Switzerland — they tried to have a computer that simulates the whole human brain. We don’t understand enough about the human brain to do that. And even if we did, as professor Ahmed said, small errors in calculations accumulate and you have completely different behavior.”
Note: Hamza explicitly referenced “professor Ahmed” — confirming Speaker 1’s surname as Ahmed.
Connection to IEIA Architecture (Full Picture)
| Speaker 2’s Concept | IEIA/MaestroHub Equivalent |
|---|---|
| Input space (raw images) | Raw PLC/SCADA tag data |
| Group action (symmetries) | ISA-95 hierarchy (Enterprise/Site/Area/Line/Cell) — the structure that organizes the data |
| Orbit (equivalent inputs) | Same measurement from different sensors/timestamps — same semantic meaning |
| Quotient space (simplified) | UNS Semantic Model (Energy Cost Mapping, Machine State OEE) — the abstraction |
| Equivariance | Dependency graph — a change at L1 (PLC) propagates predictably to L3 (MES) and L4 (SAP) |
| Data efficiency | 77 pre-built models — no schema design from scratch; abstraction already done |
| Generalization | OCP pilot schema → same models work at Safi, Khouribga, Laayoune without redesign |
Immediate IEIA Actions
1. Speaker Identification Status ✅ BOTH CONFIRMED
Speaker 1 (OT / Vanguard Center): ✅ FULLY CONFIRMED
- Ahmed Ratnani — ratnania.github.io · ahmed.ratnani@um6p.ma · LinkedIn: linkedin.com/in/ahmed-ratnani-0a77344/
- Director, UM6P Vanguard Center (Digital Twins + Cyber-Physical Systems) · CERN–UM6P team · CYSA interim director
- PhD: Strasbourg / INRIA · Post-doc: Max-Planck Plasma Physics Munich
- H-index 16, 1,006 citations · GitHub: github.com/ratnania
- → Ready to contact now
Speaker 2 (ML / ETH / College of Computing): ✅ FULLY CONFIRMED
- Hamza Keurti — hamzakeurti.github.io · hamza.keurti@um6p.ma · LinkedIn confirmed
- PhD: ETH Zürich + Max Planck (advisors: Benjamin Grewe + Bernhard Schölkopf)
- Centrale Paris + Tsinghua University MSc · ICML 2023: Homomorphism AutoEncoder
- → Ready to contact now
2. Contact Ahmed Ratnani — Vanguard Center / Digital Twins Collaboration ✅ Email confirmed
Contact: ahmed.ratnani@um6p.ma
He explicitly said: “Feel free to reach us if you want to know more about our vision.”
IEIA outreach message:
“Professor Ratnani — I watched your Math & Maroc seminar on AI and national security. You said you moved to the OT side ‘because for industry this is what we need and also for defense this is what we need.’ We are building exactly that — open-architecture OT/IIoT data infrastructure for Moroccan industrial transformation, specifically targeting OCP Group’s operations at Jorf Lasfar and Safi.
You described digital twins as living ecosystems that need co-development between academia and industry — ‘if we put them at industry level, they will be outdated; if we put them at academia, nothing will happen.’ That is the exact gap the UM6P Vanguard Center is positioned to fill, and the exact conversation we want to have.
The architecture we are deploying — Unified Namespace with semantic models following ISA-95 and CESMII standards — is precisely the data infrastructure layer that your digital twin / cyber-physical systems research builds on. We believe there is a natural joint research agenda and an industry pilot pathway at OCP.
Could we schedule 30 minutes to explore what an IEIA–Vanguard Center collaboration looks like?“
3. Contact Hamza Keurti — ML + UNS Research Bridge ✅ Email confirmed
Contact: hamza.keurti@um6p.ma
Hamza is working on the AI perception / representation layer that sits directly above IEIA’s UNS data layer. His “mental twin” framing is the academic analog of the Unified Namespace → Agentic AI thesis Walker Reynolds articulates. His Homomorphism AutoEncoder (ICML 2023) is architecturally relevant to learning representations from structured industrial action-observation pairs — exactly how a UNS node emits data.
IEIA outreach message:
“Dear Dr. Keurti,
I watched your Math & Maroc seminar on AI and representation learning. Your opening framing — ‘we maintain a mental twin of the outside world’ — is the exact thesis we are operationalizing in industrial settings.
We are building a Unified Namespace (UNS) data layer for OCP Group’s phosphate processing operations: a structured, real-time abstraction of plant data organized by ISA-95 hierarchy, semantic models, and dependency graphs — precisely the kind of quotient-space reduction you described (collapsing raw sensor space into task-relevant representations).
Your work on compositional representation learning and symmetry discovery through action-observation pairs is the AI reasoning layer that should sit above what we build. Your Homomorphism AutoEncoder architecture is directly applicable to learning structured plant state representations from UNS data streams.
Would you be open to a 30-minute conversation about where your research and our industrial data infrastructure intersect? I believe there is a genuinely novel research problem here — learning equivariant world models from industrial OT data — that could produce both academic publications and real OCP deployments.
Best regards, [IEIA contact]“
4. Math & Maroc Speaking Slot
200 student members = OCP’s future engineering workforce. IEIA should present at the next Math & Maroc seminar on “UNS Architecture and Industrial Digital Sovereignty for Morocco” — bridging both speakers’ themes (national security + AI/data foundations) with IEIA’s applied industry work.
Contact: Math & Maroc YouTube @mathmaroc1396 + Instagram + July competition (24–28 July, College of Computing, UM6P Ben Guerir).
5. CMSI + College of Computing Research Partnerships
Both colleges are new and seeking industry connections. IEIA can offer:
- Industrial data environments for student research (Digital Factory Simulator, real OCP pilot data)
- Joint thesis supervision: “UNS architecture for phosphate processing plants”
- Joint publication: “Industrial digital sovereignty in North Africa — an open architecture approach”
- Internship pipeline: CMSI / College of Computing students → IEIA OT engineers
This is a talent pipeline from UM6P’s two best AI colleges directly into IEIA’s team — before anyone else builds the relationship.
Key Quotes for IEIA Pitch Decks
“Morocco does not produce any electronic device and we can be hacked because we have now a very large attack surface.” — argument for Moroccan-owned OT infrastructure
“For industry, OT is what we need. For defense, OT is what we need.” — validates IEIA’s OT focus
“Digital twins for OCP — link production digital twins to organizational digital twins — completely simulate the company.” — IROC use case, stated from inside OCP’s own university
“Moroccan talents can help Morocco to be positioned in a very interesting role across the world.” — national talent narrative
“National security is a multisectoral challenge — it includes education, energy, health, water, agriculture.” — frames OCP phosphate infrastructure as national security asset
Obsidian Graph Links
Links from this file: IEIA · OCP Group · UM6P · IROC · MaestroHub — Intelligence & Strategic Assessment 2026 · Africa Industrial IIoT Market Intelligence 2026 · Unified Namespace · Industry 4.0 · CBAM · MES · Walker Reynolds · 4Sight Holdings · AVEVA · Inductive Automation
Source: Live seminar transcript — verbatim, timestamped (22:13 → 2:28:48), April 25, 2026. Two speakers: (1) Associate Professor at UM6P CMSI, OT focus; (2) Assistant Professor at UM6P College of Computing, ML/ETH/Max Planck. Organizer: Math & Maroc (YouTube @mathmaroc1396). Transcript captured 2026-04-26. Speaker 2 identity confirmed via Gemini + web search 2026-04-26. Action status:
- ✅ Transcript complete (22:13 → 2:28:48)
- ✅ Speaker 1 = Ahmed Ratnani — FULLY CONFIRMED. Email: ahmed.ratnani@um6p.ma. LinkedIn confirmed. Ready to contact.
- ✅ Speaker 2 = Hamza Keurti — FULLY CONFIRMED. Email: hamza.keurti@um6p.ma. LinkedIn confirmed. Ready to contact.
- ⏳ Outreach to Ahmed Ratnani — draft ready above (Action #2). Send via LinkedIn or email.
- ⏳ Outreach to Hamza Keurti — draft ready above (Action #3). Send via email.
- ⏳ Math & Maroc co-organization relationship — confirm with IEIA team internally.
- ⏳ Math & Maroc speaking slot — contact via @mathmaroc1396 (YouTube/Instagram) for next seminar or July 24–28 competition.