Use of AI (ML/DL) in Science

[Astronuc]

We have many threads on AI, which are mostly AI/LLM, e.g,. ChatGPT, Claude, etc. It is important to draw a distinction between AI/LLM and AI/ML/DL, where ML = Machine Learning and DL = Deep Learning.

AI is a broad technology; the AI/ML/DL is being developed to handle large data sets, and even seemingly disparate datasets to rapidly evaluated the data and determine the quantitative relationships in order to understand what those relationships (about the variaboles) mean.

At the Harvard & Smithsonian Center for Astronphysics, AI is being developed to evaluate data and solve problems.

AstroAI strives to bring experts in artificial intelligence together with scientists to tackle the most exciting and challenging problems in astrophysics. By facilitating interdisciplinary collaborations and drawing on the expertise of the Smithsonian, Harvard and Boston area science community, we hope to advance our understanding of the universe and drive forward technology that will revolutionize and accelerate scientific discovery at the CfA.

https://www.cfa.harvard.edu/research/astroai

There will probably be an AI/LLM component as well.

What is machine learning?​

Machine learning (ML) is a subset of AI that falls within the “limited memory” category in which the AI (machine) is able to learn and develop over time.

There are a variety of different machine learning algorithms, with the three primary types being supervised learning, unsupervised learning and reinforcement learning.

https://www.redhat.com/en/blog/what-aiml-and-why-does-it-matter-your-business

https://ai.engineering.columbia.edu/ai-vs-machine-learning/

https://en.wikipedia.org/wiki/Machine_learning

Artificial Intelligence (AI): Developing machines to mimic human intelligence and behaviour.

Machine Learning (ML): Algorithms that learn from structured data to predict outputs and discover patterns in that data.

Deep Learning (DL): Algorithms based on highly complex neural networks that mimic the way a human brain works to detect patterns in large unstructured data sets.

https://www.cengn.ca/information-centre/innovation/difference-between-ai-ml-and-dl/


The New Yorker Radio Hour had a discussion with Steven Witt about the rise of AI in conjunction with Nvidia and the development of the specialized microchips essential to the AI revolution. The first half-hour is about AI (neural networks) and Nvidia (chips using parallel computing).

https://www.wnycstudios.org/podcast...-plus-elaine-pagels-on-the-mysteries-of-jesus

Across the country, data centers that run A.I. programs are being constructed at a record pace. A large percentage of them use chips built by the tech colossus Nvidia. The company has nearly cornered the market on the hardware that runs much of A.I., and has been named the most valuable company in the world, by market capitalization. But Nvidia’s is not just a business story; it’s a story about the geopolitical and technological competition between the United States and China, about what the future will look like. In April, David Remnick spoke with Stephen Witt, who writes about technology for The New Yorker, about how Nvidia came to dominate the market, and about its co-founder and C.E.O., Jensen Huang. Witt’s book “The Thinking Machine: Jensen Huang, Nvidia, and the World’s Most Coveted Microchip” came out this year.

Witt makes the comment that neural nets (networks) developed concurrently with the advanced microchips.

Witt also mentions the first, what we would call a modern AI system, was built by Alex Krizhevsky 9in his bedroom), a computer scientist at University of Toronto, who built a system with two Nvidia gaming cards (GeForce-branded GPU cards) in 2011-2012.
https://en.wikipedia.org/wiki/Alex_Krizhevsky
https://www.cs.toronto.edu/~kriz/
https://en.wikipedia.org/wiki/AlexNet
https://en.wikipedia.org/wiki/Neural_network_(machine_learning)
https://en.wikipedia.org/wiki/Convolutional_neural_network


Update/edit: When I made this post, I didn't know we had a similar post that explains aspects of Artificial Intelligence in Video

Thread 'Artificial Intelligence in Video'

Mar 31, 2024

Behind the scenes, artificial intelligence usually makes use of what is known as a Neural Network:



In image applications, an implementation called a Convolutional Neural Network is often used:



In particular, for image super-resolution a General Adversarial Network or GAN is often used:



These form the basis of modern super-resolution:



For those who are interested in the details, see:
https://arxiv.org/abs/2204.13620

But things move on. Someone thought of using a CNN to down-scale the image first...

• bhobba
• Neural
• Replies: 1
• Forum: Computing and Technology

• 

 

[jedishrfu]

Yes, there are many places where AI/ML/DL are well-suited to filtering through reams of data.

In one case, we used an ML to recognize gunshots. It involved collecting 5000+ gunshot samples, percussion sounds, and other relevant sounds, and then training a model to recognize gunshot events.

Often, an average person thinks a gunshot sounds like an engine backfire or a bag popping, but never a gunshot unless they've been in the military, police, or been around guns and gun ranges.

We were trying to build a smart door that would close upon hearing a gunshot and lock. Sadly, we couldn't get anyone interested in the idea.

Another idea was to build a smart tablet app that would alert the teacher to a gunshot and list what she must do to secure the classroom. Those few seconds of warning could save many lives.

One student built the app and had her father take her to the gun range to test it out. She also added an attendance feature and a way to report the class location and the students present to the main office in the event of a fire, gunshot, or other life-threatening incident.

 

[Greg Bernhardt]

Great thread @Astronuc! There is deep confusion and inappropriate interchanging of AI, LLM, ML, DL, Agent, MCP etc. Good to cleanly define and use each with specific intent.

 

[Astronuc]

In one case, we used an ML to recognize gunshots. It involved collecting 5000+ gunshot samples, percussion sounds, and other relevant sounds, and then training a model to recognize gunshot events.

That's an example of AI/ML for signal analysis. I know such systems have been implemented in the field.

https://sls.eff.org/technologies/gunshot-detection

According to ShotSpotter, the largest vendor of acoustic gunshot technology, the automated system’s match between a noise and a gunshot signature is verified by human acoustic experts to confirm the sound is really gunfire, and not a car backfire, firecrackers, or other similar sounds. It’s up to people, listening on headphones, to say whether or not shots were fired.

ShotSpotter is by far the largest vendor of acoustic gunshot detection systems. Currently, 100 cities are using ShotSpotter in the United States, . . .

AI/ML could perhaps analyze an acoustic signature more rapidly.

My house is near a gun range, so we hear gun shots periodically, we hear fireworks around holidays lke 4th of July, and we hear occasional backfiring on the road; there is a difference in the acoustic signatures.

Image analysis is a another form of signal analysis. Facial recognition and license plate (or automobile) reading/recognition are examples.

There is deep confusion and inappropriate interchanging of AI, LLM, ML, DL, Agent, MCP etc.

I hear that confusion in the media in discussions about AI. AI is the encompassing field, and LLM is one subset while ML/DL is another subset, and each subset has subsets (or sub-subsets). Some form of AI is being incorporated into a variety of systems, e.g., autonomous vehicles, drones, etc.

I get lots of notifications for AI-related positions from a variety of companies, both developers and users of the AI technology.

 

[Astronuc]

MCP

I had to research that term!

The Model Context Protocol is a standard way to make information available to large language models (LLMs). Somewhat similar to the way an application programming interface (API) works, MCP offers a documented, standardized way for a computer program to integrate services from an external source. It supports agentic AI: intelligent programs that can autonomously pursue goals and take action.

MCP, essentially, allows AI programs to exceed their training. It enables them to incorporate new sources of information into their decision-making and content generation, and helps them connect to external tools.

Ref: https://www.cloudflare.com/learning/ai/what-is-model-context-protocol-mcp/

From Anthropic - https://www.anthropic.com/news/model-context-protocol

Compare that to Google's AI Overview brief summary

AI MCP refers to the Model Context Protocol, an open standard by Anthropic that acts like a "USB-C for AI," enabling Large Language Models (LLMs) to seamlessly connect and interact with external data sources, tools, and APIs, standardizing how AI agents access information and perform actions, making AI development more efficient and powerful.

 

[sbrothy]

I think this one deserves a view considering the topic:

Post in thread 'Biographies, History, Philosophy of Physics'
https://www.physicsforums.com/threads/biographies-history-philosophy-of-physics.1060020/post-7296780

 

[Astronuc]

A trio of AI methods tackles enzyme design​

https://cen.acs.org/physical-chemis...rio-AI-methods-tackles-enzyme/103/web/2025/12
RFdiffusion2, RFdiffusion3, and Riff-Diff each solve different structural problems in computational enzyme design

Naturally occurring enzymes, while powerful, catalyze only a fraction of the reactions chemists care about. That’s why scientists are eager to design new-to-nature versions that could manufacture drugs more efficiently, break down pollutants, capture carbon, or carry out entirely new forms of chemistry that biology never evolved. But doing this requires placing catalytic residues with a lot of precision—something that has been extremely hard to achieve computationally.

Three new papers now report improvements to enzyme design using diffusion models, a class of generative artificial intelligence algorithms that add and then subtract noise. The models—RFdiffusion2, RFdiffusion3, and Riff-Diff—focus on different barriers in enzyme design but all have the same outcome: computationally generated proteins that successfully carry out reactions.

Researchers led by 2024 Nobel laureate in chemistry David Baker at the Institute for Protein Design at University of Washington have developed RFdiffusion2, which addresses a long-standing problem in enzyme design: determining where in the sequence the catalytic residues should go, and how the rest of the protein can lock them into place (Nature 2025, DOI: 10.1038/s41586-025-09746-w).