mariokrenn.bsky.social
Group Leader "Artificial Scientist Lab" at Max Planck Institute for the Science of Light
Artificial Intellgence as a source of inspiration in Science.
https://mariokrenn.wordpress.com/
66 posts
738 followers
382 following
Regular Contributor
Active Commenter
comment in response to
post
comment in response to
post
Wow, great work! Congrats!
comment in response to
post
Thanks a lot Robert! 🌟
comment in response to
post
Thank you Rommie! :)
comment in response to
post
When I saw this enormous speed-up (5 orders of magnitude), i knew this is the way to go for all sort of AI-discovered physics experiments. It impacted strongly the ideas in my #ERCStarting Grant, where we will expand this idea to many other domains.
comment in response to
post
Welcome! :) Sure i added you!
comment in response to
post
A fun fact: Conventional, AI-discovered solutions are complex and it takes ages to understand what is going on. In this case, the solution was way simpler than we expected thus we initially thought it cannot be correct.
Beautifully implemented by the group of Xiaosong Ma at Nanjing University.
comment in response to
post
Instead, PyTheus exploited a superposition of the origins of photons to achieve the same goal, using entirely different resources.
For me, this changed my perspective on what is necessary to create entanglement—not because I now know what is necessary, but because I’ve realized what is not.
comment in response to
post
Curiously, the algorithm kept producing something else, something simpler, which we initially thought was incorrect.
We then realized that PyTheus’s solution actually entangles two distant particles without starting from entanglement, without BSM & even without measuring all ancillary photons.
comment in response to
post
We discovered this idea coincidentally while applying PyTheus (quantum-journal.org/papers/q-202...) to quantum protocols. As a first task, we aimed to rediscover entanglement swapping, one of the most crucial protocols in quantum networks.
comment in response to
post
the reward was the closeness of the currently designed setup to the target state. This solved a problem we couldnt solve, so we wanted to know how the machine did it. After several weeks, Sören Arlt (phd in my group) found the generalizable reason, which is described in the paper posted yesterday.
comment in response to
post
The algoeuthm has access to a physical simulator, which basically follows Schrödinger equation with some simplifications one can do for low-particle quantum systems. The program was asked to find experiments that we didnt know how to set up (highdim GHZ states, highly entangled quantum states),
comment in response to
post
This experienced has motivated a more automated approach, the automated design of generalizations, by outputting python code that stands for classes of solutions. This was a huge effort (train transformers to code python for quantum), but then interpetations are simple: arxiv.org/abs/2406.02470
comment in response to
post
We had two or three solutions of a similar kind (highdim GHZs), and as we couldnt design it by hand, we went into the details of what the machine did. It took several months (on & off) to extract the conceptual idea from the messy math, by hand and by computationally fidning simplifications.
comment in response to
post
Good question. It's certainly cleverly written by my students. The other question is more challenging, depends whether you have a human-agnostic definition of the word "clever". Will think about the useage of this word more careful.
comment in response to
post
Note, it's a probabilistic multiphoton emitter, that means not comparable with simple heralded single photons. That fact allows for using it in coherent way with other probabilistic elements (and thats how the algorithm found the solution to the highdim GHZ question)
comment in response to
post
Depends on the loss of the system, without loss, about 180Hz for 4-photon emitter simulation, with realistic loss, about 40% of that. This seems to be higher than experiments and proposals for 3-photon emitters that we found.
comment in response to
post
The algorithm has discovered a way to create effective multi-photon emitters to solve quantum questions.
Now we can use the ideas in many situations, by hand - without the need of algorithms.
It's an example how we humans can learn new physics ideas from a machine. More to come -- for sure!
comment in response to
post
Indeed we discovered that PyTheus has generated an effective experimental blueprint for generating multi-photon emitters from just pair sources, by exploiting complex multi-photon interference.
Multiphoton emitters themselves are extremely challenging to build natively.
comment in response to
post
The super-human solution are experimental designs of high-dimensional GHZ states, discovered by PyTheus (quantum-journal.org/papers/q-202...).
Soeren wanted to know HOW those solutions work. What did the machine find, what's the trick, can we generalize the idea?
comment in response to
post
Yes pls, thank you!
comment in response to
post
Sure i added you
comment in response to
post
Curious about answers for these hidden gem papers by @rommieamaro.bsky.social @qzoeholmes.bsky.social @heylgroup.bsky.social @mmbronstein.bsky.social @jobrandstetter.bsky.social and many others
comment in response to
post
It’s like a keepsake from a special time. It’s even impressive I wear the same jacket from 8 years ago. Something Jensen and I have in common 🤣