My first project
We all have a love and hate relationship with AI. Whether it's cranking out the essay due in one minute or generating a lesson plan in seconds. But have you ever stopped to think about how AI is quietly revolutionizing the entire chemical industry, from designing lifesaving drugs to predicting how molecules behave? AI is changing chemistry in ways that we've never imagined before. And that's exactly what I'm going to explore today. And as you know, I major. And welcome to my presentation on advances in chemistry with AI. So let's start out with a simple question what is AI? What is actually artificial intelligence? You know, so AI refers to computer systems which are designed to perform tasks that typically require human intelligence. And they learn from experience, recognizing patterns, making decisions, and even understanding the language. But in chemistry, AI isn't just a buzzword. It's a practical toolkit that helps scientists predict molecular properties, design new drugs plants, synthesis routes, and analyze massive reaction datasets. So there is different types of AI models. So what AI models are chemists actually using today? Now let's let's break it down. Here's a visualization showing how AI model performance has evolved over time. And there are several key model types. Number one we have neural networks for property prediction. Number two we have graph neural networks that understand molecular structures and transformers that treat molecules like a language. So here's what a neural network architecture looks like. The building block of modern chemical AI. So unlike traditional computers that follow explicit rules and that are just built off of transistors, these AI models actually learn from data. So they train a neural network on millions of molecules. So a neural network is a type of long branches of data set change. And it allows the AI to think and go in different directions. It's like a brain. And it can predict properties of compounds that it's never seen before. Using certain rules we have in chemistry. It knows all the exceptions, all the rules that we have learned so far in chemistry. Now let's trace it back to how AI in chemistry has evolved. And this didn't happen overnight. Believe it or not, it started in the ancient periods of 1980s and 1990s. We have had rule based expert systems. Basically, chemists wrote down rules and computers followed them. Simple but inflexible. By the 2000, statistical machines emerged and models learned patterns from data, but still needed humans to design the features. The real evolution came after 2012 with deep learning. Neural networks started learning features automatically, and this is when AlphaFold began changing protein and science. Now, AlphaFold is a model created by Google, and the main purpose is to predict protein structures. So today in 2025, we have foundation models that are like AlphaFold. And then we have large language models like cam DFM. Now that can reason. Those models can reason about chemistry where instead of speaking the language of proteins is speaking the language of models, and it's trying to make connections to different chemistry materials. So speaking of language models, let me explain something crucial. So insistent prompts. And you've heard of system prompts probably before you know what a prompt is. But a system prompt is something an AI model is given in the backbone. And it's kind of like its true purpose in life. So a system prompt is like an initial set of instructions given to an AI model before it interacts with the user. So think of it as like a personality programming that shapes how the response for a chemical LM, the system prompt might say always conserve mass. Follow valency rules. Carbon forms for bonds. Prioritize non-toxic reagents. Site your confidence level. So, according to the arXiv paper 2505.07027, which is about system prompt, is very critical for ensuring chemical validity. So without those, an AI might suggest reactions that violate the fundamental laws of chemistry like creating atoms out of nothing, conservation of energy, and all other aspects. Also, you have LLM augmented synthesis with the right prompts. AI models can do something very remarkable. They can plan the entire synthesis rounds automatically. So the stuff we've been doing in organic chemistry, creating multi-step an AI model can do it instantly without anyone being involved in it. So, as we know, traditional synthesis requires significant planning hours and hours of work by chemists trying to see what they have and what they need and work backwards from a target molecule. But it takes hours, even days. And out of them, augmented systems like LM. S Wind Planner from the arXiv 2505.07027 uses evolutionary search combined with LM reasoning to generate and optimize complete retro synthesis pathways in minutes. Here's a visualization of how AI plants a reaction pathway. Now let's talk about the specific breakthrough that we've been talking about over a long period of time. Chem DFM. So the Chemical Dialogue Foundation model, which is, according to the arXiv paper to 401.14818 cam DFM was trained on 34 billion tokens from over 3.8 million chemistry papers and 1400 textbooks. It was then fine tuned with 2.7 million chemical instructions. This means that the model can now think with the capacity of over a million scientists. So let's visualize a molecular structure that can DFM can analyze in real time. Now what makes chem DFM special? Why can't we use Gemini or GPT four or GPT five? So GPT four and GPT five treats chemistry as regular text. But can DFM understands smiles? Notation. This is a way that chemists write molecules as text strings, and it speaks chemistry natively. Its first language is chemistry. Now here's what's super interesting. Despite being only 13 billion parameters, which is actually small for a model. Models like ChatGPT and GPT five and Gemini are over 1.5 trillion, and even goes up to 5 trillion for parameters and parameters. Is the capacity of those neural networks how many connections that neural network can make? So can DFM actually outperforms GPT five and current Gemini platforms on chemical benchmarks? The size of the model isn't everything. This is what matters. Specification, specialization, and understanding the actual content and being trained on it. What it's what matters. And this is what research has found out. So how does cam DFM compare to the AlphaFold models we've had before. And they're so and they're actually complementary and not competitors. So you've probably heard about AlphaFold. As I said before, AlphaFold predicts what a protein looks like. You know, folding amino acids sequences into 3D structures. It revolutionized structural biology. And here's a protein structure visualization, the kind of output AlphaFold produces. Can DFM, on the other hand, focuses on chemical synthesis and reasoning. It helps design molecules that interact with proteins one predict structure and the other designs function. So AlphaFold predicts the structure and DFM designs the function. So the arXiv paper of 250.2502.11326 on IDP four actually addresses Alpha force limitations. It struggles with disordered proteins that don't have fixed structures. The field keeps pushing forward to go over this barrier. So finally, let's talk about retro synthesis. It's one of the most exciting applications of AI in chemistry. Now ratio synthesis is the art of working backwards. So you start with a complex surrogate molecule and you break it down step by step into simpler, commercially available starting materials. It's kind of similar to what we did in organic chemistry. And the breakthrough from arXiv 2510.16590 is atom anchored atoms, so these models design unique identifiers to every atom in a molecule, and they track them through the reaction. This atom level reasoning achieves over 90% accuracy in identifying where reactions should happen. And 94% accuracy in predicting the final reactions so no atoms are lost or magically created. So let me show you something interactive. This simulation demonstrates collision theory and how temperatures and concentration effects rate reaction rates. So we focus on collision theory a lot and this will help visualize that. So to summarize everything we've covered today, AI and chemistry has evolved from simple rule based systems to powerful foundation models like Cam, DFM and system prompts ensure chemical validity, LLM augmented programs automate synthesis planning, and atom anchored reasoning begins and unprecedented precision to retro synthesis. It's from generating essays to revolutionizing drug discovery. AI is proving that its potential stretches far beyond beyond what we initially imagined. And in chemistry, it's not going to replace scientists any time. It's giving them tools, utilities they can use to overcome obstacles to become something more than what we are now. Thank you for joining my exploration of AI in chemistry. And this has been Americana, and I'm pleased to show you this. Feel free to walk around this platform, and there's can be a form that you can actually use in real time. It uses a similar system prompt that chem DFM actually has, and it will help you understand the true workings behind the model. You can ask any questions related to chemistry and have fun. Thank you for watching.