With a Focus on Hydrogen sensor Application

We have crossed a dangerous threshold in industrial AI, facing a “Calibration Crisis” where 30% of sensors drift and/or are miscalibrated (per McKinsey), and many planners are giving up on physical world data. And instead, they are pivoting to Synthetic Data and Digital Twins.

The logic is seductive: If we can’t trust deployed sensors to create consistently accurate data and can’t easily collect it in the Cloud to later train our AI models, we can now use Digital Twins with Synthetic Data to simulate perfect training data instead.

It is more than a trend, as Nvidia says 75% of AI training will use synthetic (computer generated) data in 2026, not real-world data.

But that’s misleading, as the more complex applications (like Hydrogen) are more in the 15-20% range. In either case, a major trend. Let’s explore the good and the bad.

This creates a dangerous disconnect as the synthetic world and the real world get out of sync. Where do you stand on the self-audit checklist below? Let’s find out.

Note:

Do not miss at the end, “The Take away for AI Planners”, the “100% Blind” scenario, and the “Warning” paragraphs for those selecting or thinking about selecting synthetic approaches. Also, this article can be considered a continuation of the issues I presented in my prior APC article, “Is There a Sensor Calibration Crisis”. And good background for those interested in this topic.

The 7 Levels of Industrial Sensor Reality: A Self-Audit Checklist

1. The “Legacy Trap” (High Real-Time Risk)

You rely on a deployed fleet of sensors for real-time safety and efficiency. But you are facing the “McKinsey Reality”: nearly 1 in 3 of your sensors is miscalibrated and has likely drifted, with no actionable notice. You aren’t just getting a significant amount of corrupted data; you are actively flying on false instruments.

This creates immediate operational risk (false alarms and/or missed alerts) before you even think about AI. See the table below that shows this problem state called “Legacy Blindness”. Is this you or your company?

YES OR NO?

2. The “Poisoned Well” (AI Training Risk)

You want to build predictive models, but you have a historical data problem. Either you haven’t stored your operational data over time, as it occurred (The Void), or you have stored it, but it is corrupted by that same 30% calibration error (The Poison).

If you knowingly train your future AI on this drifted historical data, you are baking fiction into your foundation. And possibly adding some liability, see #7 The “Liability Void” below.

YES OR NO?

3. The “Context Gap” (Data Science Risk)

You might have deployed smart, accurate, autonomous self-calibrating sensors but you are intimidated by the IoT challenges to gather the data to the Cloud, and then there’s the Context Problem.

You have good timestamps, but you lack the “Meta-Tags” (event, status, weather conditions, maintenance codes, operator actions, …) needed to explain why a reading spiked. Without this context, even accurate data is just “noise” (not signal) to a predictive AI model. Time to step it up on context logging.

YES OR NO?

4. The “Synthetic Bypass” (Strategic Risk)

You have given up on the cost, hassle, and delays of remote, manual calibration to assure accuracy. Instead, you are joining the growing number of companies (per Nvidia) pivoting this year to Synthetic Data. You are deciding to train your models on computer-generated simulations rather than fighting the battle for clean, accurate, historical real-world data.

Nvidia and Siemens are building tools to help you simulate factory futures with “physics-level accuracy” to close the gap between ideas and reality. There’s a risky paradox here, see #5 below.

YES OR NO?

5. The “Operational Bind Spot” (The Hallucinating Twin Risk, a Paradox)

This is the most “dangerous scenario” as there are certain unmet assumptions. You have successfully used synthetic data to build a brilliant predictive model. But because you gave up on the physical sensors (Point 4), your very “Smart Brain” is now connected to “Drifting Eyes.”

Your AI knows exactly what a leak looks like “theoretically”, but your deployed sensors are potentially too mis-calibrated to see/detect it happening in real-time. A major disconnect with a false sense of security.

YES OR NO?

6. the “Edge Case” (The Hybrid Toe-Dip, a Smart Play)

You are using the smartest possible strategy: A “Hybrid” model. You rely on accurate real-world data for the 99% of normal operations (because you solved the calibration issue discussed above). You strictly limit synthetic data to the “Edge Cases”, the rare catastrophes and major disputations like transformer explosions that you can’t safely test in real life. Or wait to happen just to get the sensor readings that preceded it.

The Payoff: You get the best of both worlds. Your AI is grounded in reality for the daily grind but trained on simulation for the rare disaster. This is the less expensive and smart way to begin to use synthetic data, as a supplement, not a substitute.

YES OR NO?

7. The “Liability Void” (The Audit Risk)

This is the point that wakes up your General Counsel. If your predictive safety system fails and an accident occurs, auditors will ask: “Was the data certified in this system?”

If you answer, “We used a simulation because our real sensors were unreliable,” you are defenseless. Relying on synthetic models for safety-critical decisions creates a “Black Box” of liability. Accurate, verifiable sensor data is not just an operational asset; it is your legal insurance policy.

YES OR NO?

The logic is seductive: If we can’t trust deployed sensors to create consistently accurate data and can’t easily collect it in the Cloud to later train our AI models, we can now use Digital Twins with Synthetic data to simulate perfect training data instead.

A WARNING About Deciding to Use Synthetic Data and Digital Twins

Before you abandon the physical world for a synthetic one, recognize the hidden price tag. Building a multi-mode (1. digital/simulated vs physical/real-time), physics-accurate Digital Twin is not just a software upgrade. It is a very complex, capital intensive construction project requiring specialized software, expensive talent (generally different skills than your expensive AI team) and needs constant maintenance.

It could be cheaper, faster, and safer to simply solve the problem at the source: deploy autonomous, self-calibrating sensors at the start, those that generate in real-time (for alerts and alarms). And that also produce the consistently accurate, real-world operational data you need. Plus, that also gather it on-sensor and/or via IoT connectivity for later AI model training. Note – do not be fooled by the “auto-calibration” pitch as that’s still a manual, in-the-field effort.

Of course, doing both, solving the sensor mis-calibration issue and building a simulated digital twin, if you can afford it, would work well together. Edge cases discussed in #6 above are less expensive way to get started.

I’ll write more on this interesting but complex issue in the future.

BOTTOM LINE: Synthetic data solves the future training problem, but it cannot solve the present sensing accuracy problem. Whatever path you choose, safety and efficiency start with accurate, autonomous self-calibrating sensors that ensure the numbers on the screen match the reality in the pipes and tanks.

See this informative chart below.

The Ai Paradox: Are We Building Perfect Bains for Blind Machines?

AI Model Source (the Brain)	Real-Time Sensor Status (the Eyes)	Operational State	The Consequence
Corrupted (Bad Historical Operational Data)	Drifting (Bad Real-Time Data)	Legacy Blindness	Hight Trust / High Drag. Operators ignore alarms because they know the data is bad. Efficiency suffers, but caution is high.
Synthetic (Perfect Operational Data & Logic)	Drifting (Bad Real-Time Data)	False Security	High Trust / Extreme Risk. The “Blind Spot”. The AI confidently reports “System Normal” while a leak occurs, leading to undetected disaster.
Synthetic (Perfect Operational Data & Logic)	Drift-free, Self-Calibrating (Accurate Real-Time Data)	True Operational Resilience	High Trust / Verified Safety. The AI has perfect logic and accurate eyes. It correctly detects anomalies and predicts failures.

BONUS – Want Another Self-Audit View?

The risk Matrix: Another Way to Look at It

This table summarizes the dangerous trade-off planners are sometimes unknowingly making when they choose synthetic data without fixing the calibration issues with their sensor fleet first, those that must ensure accurate real-time data:

The Takeaway for AI Planners

As we rush toward this synthetic future, we must not confuse simulation with the situation.

• Synthetic Data is for Strategy: It teaches the system what could happen, theoretically.

• Autonomous Sensor Calibration is for Survival: It allows sensors to continuously tell the system what is actually happening.

You cannot swap one for the other. If we invest millions in the “AI Brain” but neglect the “Eyes” (sensors), we aren’t building smart factories. We are just building very confident, often blind machines.

Final Thought: The “Zero Visibility” Baseline

If you are one of the 25% of industrial sites (per McKinsey) that has not yet deployed sensors at all, do not feel relieved by the “30% drift” statistic.

The article image at the top shows that facilities with drifting sensors are 30% Blind. If you have no sensors, you are 100% Blind (except for sporadic manual inspections).

But remember: You cannot train a real-time AI on a clipboard. For the purpose of modern predictive safety and efficiency, manual data is effectively invisible (aka near useless).