Recommendation Reference Info

Hi Dr. Smith!

The programs are:

• Garcia Center (hands-on material science research) Due Wed 3/4
  • Submission link they told me to share
• YSPA (astrophysics !) Due Mon 3/16
  • Submission link: should have been sent to your email

No need to read all this! Just use whatever you find useful. And please let me know if you would like me to share anything else in particular. Thanks again Dr. Smith.

1. Getting to explore physics from 9th grade

I really first became obsessed with physics many years ago after I read Carl Sagan’s Cosmos and just naively wanted to understand what the heck the universe was made of.

I watched popsci explanations but there’d come a point where I couldn’t follow no matter how hard I tried. I just didn’t have the foundation necessary.

So 3-4 years ago I started to spend most of my time outside school first learning calc and then worked through a fundamentals textbook (happens to be the one we use in class haha), more in-depth mechanics, and now the olympiad handouts recommended by physics team.

I’m infinitely grateful when you took the time to support me since freshman year, helping me get into physics here at TJ and encouraging me to take it early.

• I’ve gotten to meet some incredible friends and got the privilege to perform PHYSICAL experiments — to really feel this discipline coming to life before me. It’s been thoroughly amazing.
• I know I don’t speak very much in class, but I really do enjoy listening to your lectures and the discussions.
  • In particular your foresight is something I could never get from textbooks, e.g. how to think about things like radians / units, what topics we need to intuit and which ones are okay to take for granted now and drill into later, how flipped charges were an unlucky mistake not because of some mysterious ultra-deep physical reason. Even just today, I really liked the $\Delta V = - I R$ way to think about circuits.
• An aside: Given all this, I wanted to give back to the TJ physics community. If there’s any way I can help, please let me know I’d love to.

2. Wonder?

I get a weird sense of awe whenever I step back and realize that every part of experience really does follow these weird symbols like $\tau$ and $E$. Most people I’ve talked to don’t feel the need to double over and catch their breath in delight whenever they deeply understand a new concept or picture atoms jiggling at near lightspeed, knowing that that is what creates their conscious experience. And they probably get a lot more done because of that.

I’m still trying to figure out whether this feeling is actually useful or not, and the true source of it (monkeys shouldn’t have been wired to do this?). But the point here is just that I really do love learning about the universe.

So although I didn’t do great on F=ma in 9th grade or 10th, the momentum is conserved.

(can’t wait to dig into quantum mechanics from knzhou and Griffiths this spring.)

3. Stubborness

I still remember that first time when I was reading Thinking Physics and the book used $F=ma$ as the solution to a problem, and for days I just felt rage.

Why?? Is $m$ the fundamental property, and $F$ and $a$ are defined based on that? But I read that $m$ is a measure of resistance to acceleration, so is $a$ more fundamental? I just could not proceed once this happened and would spend days reading stackexchange or chatting with friends (or now the experts at physics team or with Gemini). Only finally when I get a satisfying answer, ideally a mental model I can fluidly manipulate, can I move on. (or sometimes, I’d have to unwillingly concede that physics is just models of the world and not derived from axioms like math)

This same pattern keeps repeating itself, e.g. when I first encountered light (what’s a charge??), heat (isn’t entropy subjective??), and many more. It gets frustrating, though that’s the fun part.

4. Integrity

In my binding affinity research project, I tried to apply the same rigor from physics to computational predictions, which in some ML fields are notoriously number-chasing and not usable.

This means I had to be careful that my model was learning the right thing. At first (before learning about the inapplicability of many models), I used random train/test splits and iterated model designs until I reached the best performance on those benchmarks. But I later realized it was just memorizing answers this way, so I started splitting by molecule, RNA, and disease types. Performance initially fell, but the resulting model was able to demonstrate much better diversity to a broader range of applications. Excellent! Science at work.

And physics does inform biology ML models too! One of the biggest challenges was finding a way to fuse molecule and RNA information. Across several weeks, I tried everything the literature recommended (concat, standard attention, prenorm, postnorm, positional encodings, early fusion with graph neural networks, and so on). Each training run took about two days on GPU. And more than once, I’d wait 48 hours only to discover a bug at the very end. So I went back to first principles. How does binding actually happen physically? Each atom on the RNA and molecule influences each other. They’re not one-directional. So I came up with a symmetric cross-attention system where both sides attend to each other simultaneously. And then, inspired by DeepMind’s Flamingo vision model, I added tanh gating to control how much to trust each signal. What I realized was that in real binding, some parts of the RNA and molecule are totally incompatible. Standard attention forces them to interact anyway, creating noise. The gates let the model ignore incompatible regions. By actually using the physics to inform structure, the model was finally able to break through!