The following is the description of Physical AI offered up by Gemini (ChatGPT comes up with a similar description). The definition distinguishes physical AI from AI in the digital realm e.g. recommendation systems, etc.:
These and other explanations fail to paint a simple and holistic picture of Physical AI that answers the questions:
What?
Why?
How?
Therefore, I started the newsletter series by first attempting to demystify Physical AI through examples. A review of past articles leads to classification of physical AI into two distinct categories.
#1 Physical AI built on Machine Generated Data
Modern cyber-physical systems include:
Automobiles
Trains
Power generation and distribution
Water delivery
Waste management
Buildings
Hospitals
Airports
Healthcare
Agriculture equipment
These are often constructed with adaptable hardware, sensing, and communications. The systems are called cyber-physical because they are an integration of physical and cyber systems.
Automation using closed loop control is a prominent objective in these systems. Other outcomes sought are anomaly detection, diagnostics, prognostics, pattern recognition, visual analytics. These outcomes deliver business value through:
Enhanced performance
Experience (self-driving vehicle)
Increase in efficiency (energy savings)
Speed
In order to achieve these objectives, the systems have to be built with adaptability (think of this as actuators or knobs that can be adjusted) and sensing.
AI is a key horizontal technology in the physical and cyber stack of technologies that make up these systems. The objective of AI is to analyze, infer and act. A vehicle, as shown in our first article, Beyond Chatbots: AI’s Role in Physical Systems & Society, is a great example as shown in the figure below.
This is an example of Physical AI. It is built on machine (vehicle sensor) data and actuators designed into the vehicle. In the example shown below, data may be broader than the vehicle e.g. weather, road conditions, etc. Use of a broad dataset —machine and environs—is also called Broad AI.
In Physical Systems, AI or not to AI should be first question
The objective in the example of the self-driving vehicle is self-driving using closed loop control (inset block diagram in the figure). The objective—self driving with closed loop control—may not need AI at all. Therefore, the following are key questions to ask.
There is no need for AI if system behavior (the physics) is well understood, and a “plant” function (inset in the figure above) can be derived with analytical models. The use of physical science domain knowledge suffices e.g. autopilot in an aircraft.
If the system is very complex (e.g. it has many inputs, such as several or even thousands of sensors), and many outputs (actuators), AI is applied to infer and act to
detect anomalies, discern root cause, drive corrective actions
automation through closed loop control (robots, self-driving car, environmental control, etc)
💡 Note: Even in this case, AI is applied using domain fundamentals e.g. using reduced order models, running experiments, to train the model. We will cover this in future newsletters.
#2 Generative AI in Physical Systems
The next series of articles delved into the use of generative AI applied physical systems. We explored when and how to use generative AI in physical engineering workflows.
We used examples, and applied ChatGPT and Gemini, to show that:
There is no need to use generative AI if the answer lies in well-known fundamentals e.g. a well-understood case of resonance. We articulated this using a vivid example in the newsletter entitled AI or not to AI: Before you prompt, try Physics.
As engineering design becomes more complex, and a deeper understanding is sought, domain specific prompts can expedite workflows. Please refer to our newsletter entitled Prompt Engineering ≠ Not a Short Cut to Expertise.
Domain specific prompts require a framework built on fundamentals of a given field e.g. we showed an example of a framework of mechanical engineering
Generative AI can be a good companion to textbooks, particularly when learning about new developments since the textbook was published e.g. the case of Tacoma Narrows Bridge. We discussed this example in the newsletter Prompting a Bridge Collapse: Classic Textbooks vs. GenAI.
💡 Note: The application of generative AI in physical systems, using domain specific prompts, is an example of Physical AI. In future articles, we will discuss generative AI trained on institutional knowledge and private data e.g. in the case of the vehicle, proprietary vehicle details and institutional knowledge.
Classifying Physical AI given the examples
Physical AI can be classified as shown below. The reader is requested to opine, disagree, and suggest additional categories:
Use of data, domain fundamentals and AI with the objective to build autonomous systems (closed loop control), and perform diagnostics, prognostics, pattern recognition, predictive maintenance. And successful physical AI solutions result from big picture systemic integration of physical and cyber technologies. The AI methodologies are called (we will discuss these in future newsletters).
a. Statistical AI
b. Machine Learning
Use of domain specific generative AI e.g. in engineering workflows during design and operation phase.
About the Author
Former SVP, Chief Engineer, and Senior Fellow at HP, Chandrakant is a leader in AI, energy-efficient computing, and sustainability. He is an IEEE Fellow, ASME Fellow, member of the NAE, and the Silicon Valley Engineering Hall of Fame. Follow me on LinkedIn or email me at [email protected].
Want to stay ahead of the curve with insights into the newest advancements in Physical AI? Subscribe to Chandrakant’s newsletter at GenAI Works.
🚀 Boost your business with us—advertise where 10M+ AI leaders engage
🌟 Sign up for the first AI Hub in the world.
