# Artificial intelligence predicts future directions in quantum science

An episode of [Physics World Weekly Podcast](https://invisiblenetwork.ai/podcast/1400686733), 4 Sep 2025, 37 min.

Audio: https://media.blubrry.com/physicsworldweeklypodcast/content.blubrry.com/physicsworldweeklypodcast/PWWeekly4Sept2025PredictingQuantumFuture.mp3

This podcast explains what happens when 66,000 research papers are used as training data

## Transcript

[0:08] Hello, and welcome to the Physics World weekly
[0:11] podcast.
[0:11] I'm Hamish Johnston.
[0:13] In this episode,
[0:14] I'm joined by two physicists
[0:17] who are exploring
[0:18] how artificial intelligence
[0:21] can be used to predict future directions
[0:24] in quantum science and technology.
[0:27] They are
[0:28] Mario Kran, who heads the artificial
[0:31] scientist lab at Germany's
[0:33] Max Planck Institute for the Science of Light,
[0:37] and Felix
[0:38] Frohnert,
[0:39] who is doing a PhD
[0:41] on the intersection
[0:43] of quantum physics and machine learning
[0:46] at Leiden University
[0:48] in The Netherlands.
[0:50] Earlier this year, Felix,
[0:52] Mario, and colleagues
[0:54] published a paper called discovering
[0:57] emergent connections in quantum physics research
[1:01] via dynamic word embeddings.
[1:04] And they join me today to chat about
[1:07] that work
[1:08] in which they analyzed more than 66,000
[1:12] abstracts
[1:13] from the quantum literature
[1:15] to see if they could predict
[1:17] future trends in the field.
[1:27] Hi, Mario and Felix. Welcome to the podcast.
[1:30] Hi, Mitch. Very nice, to see you. Glad
[1:33] to be here.
[1:34] So, Mario,
[1:36] can we start off by sort of delving
[1:38] into the history of this?
[1:40] Apparently, in the past, people have tried to
[1:43] use,
[1:44] data from the scientific literature
[1:47] to predict the evolution
[1:49] of a research field. Can you talk a
[1:53] bit about some of that early work and,
[1:56] and how successful it was?
[1:58] Yes.
[2:00] There have been a lot of, work in
[2:02] the past where people try to use this
[2:04] enormous amount of data and see how
[2:08] researchers work and how researchers could work. So
[2:12] looking at the strategy of how,
[2:15] research is done. This has interestingly,
[2:17] mainly been done by computational
[2:19] sociologists
[2:21] who look
[2:22] at science as a whole and analyze
[2:25] what
[2:25] strategies that people use and how we could
[2:28] accelerate those strategies.
[2:30] And maybe one of the most exciting work
[2:32] here in this field that at least,
[2:35] motivated me to look into this was a
[2:37] paper from
[2:39] 02/2015,
[2:41] led by,
[2:43] Chicago
[2:44] computational sociologist,
[2:45] James Evans.
[2:46] And that paper, it was published in PNAS.
[2:49] It was called choosing experiments
[2:52] to accelerate
[2:53] collective discovery.
[2:55] And, therefore, the first time, I've seen how
[2:58] people use millions of scientific papers and patents
[3:02] to analyze what other concrete strategies that humans
[3:06] did,
[3:08] showing that those strategies are at a collective
[3:11] scale actually quite,
[3:15] improveable,
[3:15] let's say, like that. And then they did
[3:17] a lot of computer simulation to come up
[3:19] with potential different strategies.
[3:22] And they have pointed out that if you
[3:24] would follow if the collective would follow different
[3:27] strategies, you could theoretically,
[3:30] come to much faster,
[3:33] progress in science.
[3:34] But this was in 02/2015,
[3:36] and,
[3:37] this motivated
[3:39] me at the time that was still during
[3:40] my PhD to also try the same thing
[3:43] in my own field in quantum physics. And
[3:45] then we also,
[3:47] tried to use hundreds of thousands of papers
[3:51] to predict,
[3:54] what scientists will do in the field of
[3:56] quantum physics. And then we use the same
[3:58] techniques that James Evans, teams have used. Knowledge
[4:02] graphs where you have concepts that are the
[4:04] notes and edges
[4:06] that are formed when two concepts
[4:08] are connected,
[4:11] by a single paper. So you see somehow
[4:13] the evolution
[4:14] of,
[4:15] of what scientists
[4:17] did in the past.
[4:20] And that a number of,
[4:22] there are a number of follow-up papers came.
[4:24] So this one was published in 2020 in
[4:26] PNAS, and then we did the AI competition,
[4:29] but all was based on the same idea
[4:32] of using this knowledge graphs.
[4:34] And then,
[4:36] I met,
[4:37] Felix,
[4:39] and Evert,
[4:41] and they had a very different idea of
[4:43] how one could
[4:44] integrate
[4:45] scientific knowledge and especially the dynamics,
[4:48] and that is what this specific paper is
[4:51] about.
[4:52] I see. Okay. And and I I suppose
[4:55] it's well, is it obvious? I mean, it
[4:57] makes sense to me that you've got a
[4:58] huge amount of data, and that's where machine
[5:01] learning, artificial intelligence
[5:03] can really help. And and I think in
[5:05] the paper that you've published recently,
[5:08] you've looked at how machine learning can be
[5:10] used to look for
[5:11] emergent connections between subfields
[5:14] in quantum physics with the ultimate goal of
[5:18] forecasting
[5:19] the future direction of research.
[5:22] Can you explain what a subfield is and
[5:25] and perhaps give an example
[5:27] of an emergent
[5:29] connection. I think may maybe a lot of
[5:30] our listeners would know what a subfield is,
[5:32] but
[5:33] the emergent connection bit
[5:35] might be something new to them.
[5:38] Felix, could you, could you address that?
[5:42] Yeah. Sure. Of course. So for us,
[5:45] subfields in
[5:47] quantum physics within kind of this broader category
[5:50] would be something like,
[5:51] single photon quantum optics. Different sublet would be
[5:55] gravitational wave physics
[5:56] or topics related to, I don't know, fault
[5:59] tolerant quantum algorithms.
[6:01] All of these would fall under the umbrella
[6:03] term of quantum physics, but in of itself
[6:06] deal with very kinds of different questions and
[6:09] techniques.
[6:10] While quantum optics might focus on questions like
[6:13] how to generate,
[6:14] interfere, or detect single photons,
[6:17] topics like fault tolerant quantum algorithms
[6:19] might deal with questions about error correction
[6:22] or resource estimation,
[6:24] which, of course, fall under this umbrella term
[6:26] of, quantum physics but are conceptually quite
[6:30] distinct. But what we have seen in the
[6:32] past
[6:33] is that despite these differences, there can be
[6:36] fruitful overlaps
[6:38] between different
[6:39] subfields
[6:40] in general in science.
[6:42] One example that is quite prominent
[6:45] in the,
[6:47] kind of physics community has been, I think,
[6:49] the application of, machine learning to the study
[6:52] of quantum
[6:53] phase transitions,
[6:55] which is a connection that has only been
[6:56] made in around,
[6:58] 2017,
[7:00] I think.
[7:00] Before that, these two
[7:03] fields
[7:04] were kind of completely distant. Machine learning at
[7:07] that time was mostly used for things like
[7:09] image classification or regression,
[7:11] and quantum phase transitions were studied with kind
[7:14] of very
[7:16] physics specific
[7:17] tools.
[7:18] But then in around, 2017,
[7:20] researchers realized that you could use measurements of,
[7:24] quantum systems to kind of treat them as
[7:26] structured data similar to images, making it possible
[7:29] to kind of combine these two techniques or
[7:32] apply machine learning methods from computer vision to
[7:35] physics problems.
[7:36] And this kind of quite simple connection but
[7:40] powerful idea,
[7:41] Combining these two distinct ideas
[7:44] has led to many fruitful results, basically. And
[7:47] the goal
[7:48] for
[7:49] this research project,
[7:51] Mario and I did, is to try to
[7:53] build or kind of improve a data driven
[7:55] method methods that helps to assess or even
[7:58] forecast
[7:59] ideas of kind of that type, basically.
[8:02] I see. And so, I mean, one thing
[8:04] that
[8:05] I can think of in, you know, in
[8:07] in experimental physics, for example, And I'd like
[8:10] to see, I mean, have I got the
[8:12] right idea here when I'm thinking about emergent
[8:14] connections?
[8:15] You know, for for example, you've got superconducting
[8:17] circuits.
[8:18] And, you know, in the past, superconducting
[8:20] circuits were were used to make
[8:24] very sensitive
[8:25] detectors of magnetic fields, weren't they, squids, which
[8:28] were used in medical physics and and and
[8:30] other research. But now
[8:32] superconducting circuits are used as qubits
[8:35] in quantum
[8:37] computing. So
[8:38] is that the sort of thing that you're
[8:39] looking at? You know, the the use of
[8:43] of of one quantum
[8:45] technology suddenly or well, probably not suddenly, but
[8:48] slowly
[8:49] becomes,
[8:51] useful in another area of physics? So moving
[8:54] from
[8:55] sensing
[8:56] to quantum computing. Is that, I mean, are
[8:59] those the sort of things that you're trying
[9:01] to predict in the future, or am I
[9:03] getting that wrong?
[9:06] Yes. This would be one of the examples.
[9:09] There are a number of other
[9:10] examples that are,
[9:13] where you would, as a human, potentially not
[9:16] see the connection,
[9:18] But just by having
[9:20] access to hundreds of thousands or millions of
[9:22] papers, the machine sees that there is some
[9:25] underlying connection.
[9:27] And then it can,
[9:29] say, okay. Very likely, this will be connected
[9:32] because I've seen this underlying connection before, and
[9:35] that always leads to connections.
[9:38] Yeah.
[9:40] I see. And, I mean, obviously, you know,
[9:42] being able to predict the future is very
[9:44] important.
[9:46] So who, I mean, who would benefit from
[9:49] this? I mean, it are you thinking of
[9:51] scientists who are trying to plan their career,
[9:54] or are you thinking of,
[9:56] research funders,
[9:58] looking at, you know, new areas to put
[10:00] money into or companies
[10:02] trying to work out what, you know, new
[10:05] technologies will emerge in the next little while.
[10:08] I mean, it sounds like it could benefit
[10:10] a lot of people.
[10:13] So I think for both,
[10:15] Mario and me, the motivation in the beginning
[10:18] was actually just a bit of, self interest
[10:22] at first. Because as researchers, I think one
[10:25] of the hardest tasks is to consistently come
[10:27] up with good and fruitful
[10:29] research ideas, either for ourselves
[10:33] or for our students.
[10:34] Because in order to kind of come up
[10:37] with many good research ideas, you need to
[10:39] have a mix of, like, good intuition about
[10:42] the topic, some experience, and sometimes even just
[10:44] a bit of luck to come up with,
[10:46] good ideas.
[10:47] And the hope for both of us was
[10:49] to kind of use such a model as
[10:51] sort of like an artificial
[10:52] muse in the sense that we can have
[10:54] a data driven method that helps us to
[10:57] uncover or assess combination of research topics that
[11:00] no one has considered before, where there's, like,
[11:03] no benchmark of, combining these two things
[11:07] together.
[11:08] So who would benefit from that? Basically,
[11:10] anyone working in science who regularly needs to
[11:13] generate new research ideas or kind of project,
[11:17] project topics.
[11:18] Yeah. Good benefit from this tool, I think.
[11:22] I think the biggest benefit could come from
[11:25] collaborations,
[11:27] suggesting collaborations
[11:29] for
[11:30] researchers that are not in the same field,
[11:32] potentially that are in very far away fields.
[11:35] For instance, let's say I have
[11:37] a friend who works
[11:39] in astrophysics. I know nothing about astrophysics.
[11:42] Now I want to work together.
[11:44] I would
[11:45] potentially not even know where to start.
[11:50] I have my techniques. They have their techniques.
[11:53] I have no idea,
[11:55] what is their connection. But the machine has
[11:57] seen hundreds of thousands or millions of papers,
[12:00] and they could see some underlying
[12:03] connection that might,
[12:05] be possible in the future that has not
[12:07] been possible,
[12:08] or that has not been done so far.
[12:11] So I see a lot of potential for
[12:12] those
[12:13] very surprising,
[12:15] very distant
[12:18] collaborations.
[12:19] That would be really great.
[12:21] I see. And I wanted to ask you
[12:24] about,
[12:25] the data that you use to train
[12:27] your system.
[12:29] I mean, I'm guessing you that you're using
[12:31] the titles
[12:32] of, peer reviewed papers. You're using the abstracts.
[12:35] Are you also looking at,
[12:39] references that are made in papers
[12:41] in order to to make connections? What what
[12:45] sort of input do you have?
[12:48] It's actually quite quite simple. So we use
[12:50] publicly available preprints of quantum physics papers from
[12:54] the archive.
[12:56] In general, researchers
[12:57] submitted their manuscripts to,
[13:00] websites like the archive before formal, formal publication,
[13:04] which makes the consent quite easily accessible,
[13:07] for us. And in particular, we focus on
[13:10] the abstracts of each preprint in this quantum
[13:14] physics,
[13:15] subsection.
[13:16] The idea behind that is that the abstract
[13:19] provides,
[13:21] provides a condensed summary of the kind of
[13:23] core ideas of the paper, which typically also
[13:26] includes kind of one or two sentences about
[13:29] the background,
[13:30] some information about the main results, and then
[13:32] also some information about the the impact,
[13:35] maybe even,
[13:37] about that paper.
[13:38] And only focusing on the abstract
[13:41] and not, for instance, on the kind of
[13:43] main figures in the paper simplifies the data
[13:46] quite a bit because, well, typically, in abstracts,
[13:49] you don't need to deal with,
[13:51] formulas or citations or references.
[13:54] And it also kind of, takes away all
[13:56] the guesswork with things like, figures, for for
[13:59] instance. Because what abstracts are typically text based,
[14:03] which means it's,
[14:05] much easier to kind of extract the information
[14:07] from that for the machine learning model.
[14:11] I see. And and is it possible to
[14:13] explain in simple terms
[14:16] how the the machine learning is used to
[14:19] process these data? I mean, you know, sort
[14:21] of keep in mind that you're you're speaking
[14:23] to physicists
[14:25] rather than
[14:26] computer scientists. Is there
[14:28] is there an easy way to describe
[14:31] how it's done? Or may maybe there's not
[14:34] an easy way. But, I mean, you know,
[14:36] how would you explain it to a colleague
[14:38] who's a physicist rather than a computer scientist?
[14:42] For sure. So in the machine learning pipeline,
[14:46] there are three main steps. There is a
[14:49] data collection phase, there is the words embedding
[14:52] phase, and a training and prediction phase.
[14:55] The first step, data collection, kind of connects
[14:58] to the previous question you asked,
[15:01] where we kind of take a bunch of,
[15:04] quantum physics publications, look at the abstract,
[15:08] and then from each abstract, identify
[15:11] key scientific concepts we are interested in tracking
[15:14] over time, which in our case are, well,
[15:17] quantum physics concepts, things like,
[15:19] entanglement,
[15:21] quantum circuits, or phase transition.
[15:24] The second step,
[15:26] the
[15:27] word embedding is then the kind of key
[15:30] idea of this paper. Because in order to
[15:34] use machine learning to
[15:36] connect
[15:37] these concepts,
[15:39] You need to be able to translate
[15:41] these words
[15:42] into numbers that a machine learning model can
[15:45] understand and process that somehow encodes their meaning
[15:49] and
[15:49] usage.
[15:51] And, yeah, kind of simple ideas like just
[15:54] giving each word you're interested in a unique
[15:57] number, of course, wouldn't really work because this
[15:59] would throw away all the contextual information
[16:02] about how similar or dissimilar certain words are
[16:05] or how they are used,
[16:07] in the context of these abstracts. So what
[16:10] we do instead is a technique called word
[16:13] embeddings, which basically turns each word into the
[16:16] abstract,
[16:17] into
[16:18] a high dimensional
[16:20] vector, which somehow encodes its meaning or how
[16:23] it is used.
[16:25] And then you kind of use these,
[16:29] vectors for the machine learning model. I tried
[16:32] to come up with a simple example to
[16:34] kind of highlight
[16:36] how this could look like. And for this,
[16:38] I have, like, two sentences.
[16:40] So the sun is hot
[16:42] and the moon is cold, which you could
[16:45] try to encode with such a word embedding.
[16:48] And in the end, what you would look
[16:50] for in the embeddings
[16:52] are then kind of a representation of how
[16:54] these individual words are used in these two
[16:56] sentences. So for instance, on some kind of
[16:58] axis in the embedding, the words sun and
[17:01] moon might
[17:02] share one dimension because they're used as the
[17:05] subjects in the sentences.
[17:07] The words hot and cold might then lie
[17:09] might then lie on different axis because they
[17:12] kind of represent the temperature
[17:15] of an object.
[17:17] And the idea is then that somehow the
[17:19] embedding also encodes
[17:21] how these different words relate to each other.
[17:24] So for instance, if you would compute the
[17:26] overlap between the vector
[17:28] of the word sun and hot, it should,
[17:31] if the embedding is well trained, be larger
[17:34] as then the embedding between sun and cold
[17:36] because sun and hot are kind of more
[17:38] closely aligned.
[17:40] Yeah. This is kind of the the general
[17:42] idea of how we want to encode the
[17:44] underlying information about the physics
[17:46] into these
[17:48] vectors.
[17:49] And then the only thing that this,
[17:52] machine learning model that we then kind of
[17:54] train to do the predictions in the end
[17:56] does is look at,
[17:58] two of these vectors and then try to
[18:01] predict whether these two vectors
[18:03] will appear in the same abstract
[18:05] in a given,
[18:07] time frame.
[18:09] What we do in the paper is we
[18:10] define
[18:11] a training window, so for instance, from
[18:14] 2000 to 02/2010,
[18:16] and look at which concept pairs were not
[18:19] linked yet in this tracking window. And what
[18:22] the model, tries to do is then to
[18:24] predict which of these unconnected pairs are likely
[18:28] to appear together in future abstracts
[18:31] within a time window of three years, for
[18:33] instance.
[18:34] And then the idea is that if you
[18:36] want some kind of use this machine learning
[18:38] model for inference to make kind of further
[18:41] predictions, you basically just shift the time window
[18:44] from kind of 2,000 to 2,010. You could
[18:47] also use it to predict from 2,010 to
[18:49] 2,020
[18:51] or even, further. That's kind of the the
[18:53] general idea.
[18:54] I would like to mention
[18:56] that
[18:57] what Felix just explained is the
[18:59] main difference
[19:01] from our work to what has been done
[19:02] before.
[19:03] So before I close what I mentioned in
[19:05] the beginning, the computational sociologists
[19:08] and also in my own previous
[19:10] work, we have built up knowledge graphs. And
[19:13] knowledge graphs,
[19:15] have concepts that are the vertices of the
[19:18] graph, and edges are drawn when,
[19:21] two concepts are mentioned in the paper.
[19:25] Now
[19:26] when I just use the concepts,
[19:28] I throw away everything about the
[19:32] context of that word.
[19:34] That's exactly what Felix mentioned.
[19:36] Basically, I just define
[19:38] the word as a number and remove all
[19:40] of the other information.
[19:42] But now with the technique that, Felix just
[19:44] explained that we published in this paper, we
[19:47] get much more context of the words
[19:50] by using this more modern,
[19:52] word embedding instead of knowledge graphs. That comes
[19:55] with extra technical,
[19:58] challenges that we had to,
[20:00] solve,
[20:01] But then you can get much more context
[20:04] of the word, of the phrases itself,
[20:07] and, hopefully,
[20:08] then get better predictions.
[20:11] I see. And you use data that went
[20:14] back to 1994.
[20:16] So that's, you know, we're looking more
[20:18] than thirty years ago.
[20:21] I mean, if you sort of run your
[20:23] system using older data,
[20:25] do do you find that it predicts things
[20:28] that actually happened
[20:30] a decade or two later?
[20:33] The short answer to that would be yes.
[20:37] Well, congratulations.
[20:40] And there are basically
[20:42] two different types of predictions our model can
[20:45] make. Our model can make validation predictions
[20:48] where we can still
[20:49] check the outcome and truly
[20:52] new predictions
[20:53] where we don't entirely know the outcome yet.
[20:56] As you mentioned, our model was trained on
[20:58] data from nineteen nineteen four till,
[21:01] 2017,
[21:02] in our case, where the model was kind
[21:04] of used to predict,
[21:07] contact connections within the next three years, so
[21:09] until 2019.
[21:11] And the validation predictions we did
[21:14] were then
[21:16] by, kind of shift the time window, as
[21:18] I explained for the, previous question, where we
[21:20] tried to predict
[21:22] or could yeah. Validation predictions,
[21:25] from 2020
[21:26] till 2023
[21:28] where there still is real world data that
[21:31] we can test against.
[21:33] And our model managed to kind of discover
[21:36] a couple of emergent research directions there, which
[21:39] includes ideas like using tensor network methods to
[21:42] simulate local quantum circuits, so quantum circuits with
[21:45] only minimal entanglement,
[21:47] or ideas of using machine learning to optimize
[21:50] the structure of,
[21:53] quantum circuits.
[21:54] And these kind of, two examples and a
[21:56] few more are things we discuss in the
[21:58] paper in detail, basically.
[22:01] The question about truly
[22:04] new predictions
[22:05] is slightly more complicated,
[22:08] I would say.
[22:10] One
[22:11] idea
[22:12] that I need to discuss for further first
[22:14] is the idea of,
[22:16] calibration
[22:17] of the,
[22:18] prediction model. Because
[22:21] in the end, what you are generally interested
[22:23] in is to have kind of a
[22:25] small subsets of predictions
[22:28] where the model is super subbed that,
[22:31] these predictions
[22:32] are actually,
[22:34] correct. So one thing that you need to
[22:36] check first is
[22:37] is the confidence
[22:39] of your machine learning model aligns to the
[22:42] probability
[22:43] of being right or false in this prediction?
[22:46] This is kind of one section in the
[22:48] paper where we analyze,
[22:49] like, how our model performs,
[22:52] in this aspect. And it turns out that
[22:54] this actually works kind of well. So if
[22:56] we look at the output probability of the
[22:58] model, this is aligned quite well with the
[23:00] probability of the model being false or correct.
[23:03] So what you can do then is basically
[23:06] predict
[23:07] many, many concepts, like, tens of thousands of
[23:09] concepts,
[23:10] and then rank them by their probability of
[23:14] actually occurring
[23:15] in the future.
[23:17] And we we did that and found a
[23:20] subset of predictions that have
[23:22] over 99.9%
[23:24] probability
[23:25] of occurring. And this was very interesting because
[23:27] most of these concepts
[23:29] seem to be related to single photon quantum
[23:32] optics.
[23:33] So things that are related to concepts like
[23:36] photon detector and spatial modes
[23:39] or concepts like polarization entanglements
[23:41] and,
[23:42] single photon,
[23:44] single photon source.
[23:46] And this was actually something I had a
[23:48] chat about with with a quantum opt optics
[23:51] professor from Leiden University, where I'm doing my
[23:53] PhD.
[23:54] And all of these predictions actually somehow seem
[23:57] to make
[23:58] sense, but are also already
[24:01] known. So the catch with that is basically
[24:03] that some of the publications
[24:06] that were in the kind of subset of
[24:08] quantum physics,
[24:10] abstracts that we used
[24:12] mentioned these concepts. But, of course, we did
[24:14] not specifically
[24:15] train our model on optics papers,
[24:19] which means it's kind of solve these connections,
[24:22] make these connections,
[24:23] but it never saw the actual kind of
[24:26] papers
[24:27] where these concepts are explicitly mentioned. So there
[24:29] are kind of
[24:30] two different interpretations I can make on that.
[24:33] Kind of one optimistic interpretation would be that
[24:36] our machine learning model kind of has discovered
[24:38] the concept of single photon quantum optics on
[24:41] its own,
[24:42] and the kind of slightly more conservative,
[24:45] interpretation would be that in order to get
[24:47] very meaningful predictions in the future,
[24:50] one would need to train
[24:52] our model in kind of all physics applications
[24:54] to rule out these
[24:57] kind of trivial or simple predictions
[25:00] that might have not occurred together in the
[25:02] training set, but
[25:04] might be, like, somewhere out there on the
[25:06] Internet, basically.
[25:08] I see. And is that is that something
[25:10] that's possible,
[25:12] you know, to expand the training
[25:14] to include all
[25:16] physics papers? I mean, I mean, first of
[25:19] all, you know, would you have the computing
[25:21] resources to do that, and and would that
[25:24] be useful,
[25:26] in terms of making predictions?
[25:28] Maybe you wanna speak about that, Mario?
[25:31] Yes.
[25:32] So this is,
[25:34] definitely possible. Actually, we had, different status where
[25:37] we not only used,
[25:39] quantum physics as
[25:41] input data, but also,
[25:45] entire physics. Or in one study, we even
[25:48] used all published
[25:50] papers,
[25:51] that exist.
[25:53] I just looked it up 58,000,000
[25:56] research papers. So that is possible to take
[25:59] much, much larger,
[26:02] subsets
[26:03] of,
[26:04] science.
[26:06] Of course, then
[26:07] using this data will become
[26:09] significantly
[26:10] more computationally expensive
[26:12] so that the process will be more expensive.
[26:15] But,
[26:16] if we really want to go into predicting
[26:18] novel
[26:20] research directions, then we have to do this.
[26:22] So I think in our paper, we showed
[26:24] it as this huge potential.
[26:26] And now
[26:27] we could go into,
[26:29] thinking about using it for whole physics or
[26:32] for whole natural science or just whatever
[26:35] for all science.
[26:37] Right. Okay. And is that, I mean, is
[26:39] that something that you and your
[26:41] collaborators would be interested in? Or do or
[26:44] do you think that,
[26:46] that, you know, may maybe maybe people working
[26:49] or studying other fields, people who study
[26:53] how biology
[26:54] evolves, for example, would, would join with you?
[26:57] I mean, do do do you see this?
[26:59] I mean, is this a new field
[27:01] or a growing field that,
[27:04] that you think you've made a significant
[27:06] contribution to,
[27:08] and, you know, other people will follow on
[27:10] from you?
[27:12] There
[27:13] there are, there are a lot of groups
[27:15] that look into
[27:17] this direction or related direction. It's a huge,
[27:21] field to,
[27:23] to try to come up with
[27:25] computer inspired new ideas in a, data driven
[27:29] way.
[27:30] We show one method how that is done.
[27:32] There are many other methods.
[27:35] Now the question is,
[27:37] is
[27:38] predicting what scientists will do,
[27:41] is this the one thing you need to
[27:43] discover new ideas or are there other things?
[27:46] For instance, you can think about, can I
[27:48] predict
[27:49] impactful
[27:50] research directions?
[27:52] And just a few months ago, we have
[27:54] also published in machine learning science and technology
[27:57] and,
[27:58] IOP,
[27:59] journal,
[28:00] a paper where we showed,
[28:02] that was, led by,
[28:05] my former post doc, Shumei Gu, who is
[28:07] also coauthor
[28:09] of the paper with Felix,
[28:11] that we showed that you can actually predict
[28:14] also,
[28:15] impactful research directions.
[28:18] So let's say you have two concepts
[28:20] that have never been started before together.
[28:24] Now the question is,
[28:27] will there be papers
[28:28] that started it? And the next question is,
[28:31] will those paper that studied this
[28:34] topics that have never been studied before, will
[28:36] they be impactful?
[28:38] Will they,
[28:39] create a lot of citations?
[28:41] And very surprisingly
[28:42] to us, this is also predictable.
[28:46] So now we can also predict in some
[28:48] way what future research directions could be impactful.
[28:54] Of course, what Felix has done, what is
[28:56] new, and what could, could be done and
[28:58] what might be done. And then I think
[29:00] the final question is what research directions
[29:03] are interesting
[29:05] for humans?
[29:09] For that,
[29:11] one could do,
[29:12] large scale evaluations,
[29:15] using humans.
[29:16] We have started,
[29:18] doing this, and I think there are thousand
[29:20] different things one can do.
[29:23] One thing I've seen
[29:25] in this whole,
[29:27] this whole range of work is the humans
[29:30] are exceptionally good in coming up with ideas,
[29:33] and it's really not clear where those ideas
[29:35] come from.
[29:36] I think there's a lot of a lot
[29:38] of things that we need to understand.
[29:40] From the human perspective, why are human create
[29:42] scientists? Where are those ideas coming from? That
[29:45] we can artificially recreate this.
[29:48] I see.
[29:48] And and what about, you know, this specific
[29:51] project,
[29:52] that that you, Felix, and, Mario are working
[29:56] on?
[29:57] Are you are are you following up this
[29:59] research? Are are you planning on improving
[30:02] your machine learning
[30:04] technique,
[30:06] or maybe applying it to more data? I
[30:08] suppose well, you have applied it to more
[30:10] data. What what what's next for you? Or
[30:12] are you moving on to something else?
[30:15] So there are a couple of kind of,
[30:18] obvious follow-up research directions that,
[30:21] one could pursue.
[30:23] There are,
[30:24] a couple of ideas about how to improve
[30:27] the embedding that we came up with,
[30:30] even further.
[30:31] One of these ideas would be that
[30:34] the kind of main application of our
[30:37] embedding was to kind of replace the information
[30:39] in the knowledge graph as Mavi
[30:41] mentioned earlier. But
[30:43] you don't necessarily
[30:45] need to view it in this kind of
[30:46] binary setting of either only using the knowledge
[30:49] graph or or embedding. You could also come
[30:51] up with some kind of combined representation of
[30:54] both that,
[30:56] takes kind of as much information as possible
[30:58] and then just kind of gives it to
[30:59] the machine learning model and lets it do
[31:01] its thing,
[31:02] basically.
[31:04] The other,
[31:05] obvious direction would be to,
[31:08] apply our methods
[31:10] to other fields.
[31:12] Like, on the archive alone, there are
[31:15] more than 15 other primary categories beyond quantum
[31:18] physics that we could try to use this
[31:20] method on where the data is basically
[31:23] publicly available and on the Internet ready to
[31:25] use. This includes topics like, mathematics, computer science,
[31:29] or economics or biology where we could also
[31:32] try our method on and see if we
[31:34] can find emergent research connections in these fields.
[31:40] And Mario, what about you? What are you
[31:41] looking forward to in the future?
[31:44] Yeah. So beyond things like predicting,
[31:48] more general properties like cetaceans
[31:50] and,
[31:52] maybe surprise and so on,
[31:55] one
[31:56] could one could probably make big steps into
[31:59] improving their AI models themselves
[32:01] by performing AI competitions.
[32:04] So I think the data that Felix,
[32:07] has is very well equipped for that, where
[32:10] you would have a pool of prize money
[32:11] and then a hidden dataset, and participants need
[32:15] to provide
[32:17] the best
[32:18] models that they can come up with to
[32:21] predict
[32:22] for predict, for instance, the future of, quantum
[32:25] physics.
[32:27] And then test in that case, we could
[32:30] compete,
[32:31] techniques like knowledge craft, the techniques that we
[32:34] have published in these papers,
[32:37] with many, many other techniques that other people
[32:39] come up with. And that is, I've seen
[32:42] this, now one time,
[32:45] very clearly that this is very useful. There
[32:47] are many, many diverse techniques that usually
[32:50] be in such competition,
[32:52] techniques that you alone would never think about.
[32:55] It's
[32:57] significantly more technical,
[33:00] ideas, significantly
[33:01] simpler ideas that you would not even believe
[33:03] that they might work.
[33:05] So that would be really great. And I
[33:07] think this
[33:08] type of data that we,
[33:10] prepared here would suit very well for such
[33:13] AI competitions.
[33:15] I see. And and Mario,
[33:17] you you mentioned the journal, Machine Learning
[33:20] Science and Technology.
[33:23] Can you and you're on the editorial board
[33:25] of that journal. Can
[33:27] you just give us a little bit of
[33:29] a flavor of what sort of physics related
[33:32] research the journal publishes?
[33:35] Yes. So this is
[33:38] probably one of the most important, if not
[33:40] the most important,
[33:41] journal in that
[33:43] field. Because
[33:45] machine learning research,
[33:46] getting it published in physics journals,
[33:50] can be
[33:51] quite challenging because,
[33:53] physicists are looking at different things that we
[33:56] might look at. Also publishing
[33:59] those,
[34:01] techniques
[34:01] in AI,
[34:05] conferences might also be very challenging because
[34:08] those people also look at very different things.
[34:11] So there was before this channel, there was
[34:13] really a gap where it's not clear how
[34:15] would you even publish such works. And now
[34:17] I think,
[34:19] MLST became
[34:20] the
[34:21] the two go
[34:23] address
[34:23] for exactly,
[34:25] such works.
[34:26] And then a lot of works on how
[34:29] you would,
[34:31] use new
[34:32] ideas for machine learning in
[34:34] physics and chemistry and biology.
[34:37] Our work is a little bit, outlier because
[34:39] it even goes a step further to kind
[34:41] of meet the science, but you see that
[34:44] MLC is, quite open in,
[34:46] disrespect.
[34:48] So,
[34:49] that's why I'm super happy that the channel
[34:51] exists and also super happy that they invited
[34:53] me to join the, editorial board.
[34:57] Well, that's great. Thanks. Thanks, Felix, and Mario
[35:00] as well. Thanks for coming on the podcast
[35:02] and talking about your research. It's,
[35:05] it's a fascinating
[35:06] application of, machine learning,
[35:09] and artificial intelligence.
[35:10] And, yeah, we hope to, to learn more
[35:14] about it from you and your colleagues. Thank
[35:16] you.
[35:17] Thank you very much. Thank you. Bye.
[35:27] That was Felix Frohnert of the University of
[35:30] Leiden and Mario Krenn of the Max Planck
[35:34] Institute for the Science of Light.
[35:37] Thanks to both of them for a fascinating
[35:39] discussion.
[35:41] You can find their open access paper in
[35:44] the journal,
[35:45] Machine Learning,
[35:46] Science and Technology,
[35:48] which can be found on the IOP Science
[35:51] website.
[35:52] Just look for the title,
[35:54] Discovering
[35:55] Emergent Connections
[35:56] in Quantum Physics Research
[35:59] via Dynamic Word Embeddings.
[36:02] I'm afraid that's all the time we have
[36:04] for this week's podcast.
[36:06] I'll sign off with a thanks to our
[36:08] producer, Fred Ailes, and an invitation for you
[36:12] to join us next week.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.