Artificial intelligence is a hot topic, but should I pursue it if I'm interested in Matter Modeling?

Question

Since my previous question seems to have a large number of views, I would like to ask my central question.

I am a student now and should decide my major to research soon. AI/ML is a very hot research field and I am very interested in it, but I have some doubts before I study AI/ML.

For example, a paper that my professor told me is a famous research and has about 300 citations in two years. However, I found the GitHub issue of this paper, and the high-accurate ML model (called DimeNet, which is a graph neural network model) can not predict the energy of hydrogen molecule at all (the error is over 100 kcal/mol!). The reason, according to the author of this paper, is that the training dataset (this paper used the benchmark QM9) does not include the hydrogen molecule. (It feels like an AI can not translate the sentence "I like cats" into German because "Ich liebe Katzen (I like cats)" does not exist in the training dataset.)

I have examined the QM9 benchmark dataset and display some molecules and their energies as follows.

As shown in this figure, many data samples have very similar molecular structures and energies, so I believe that even children can easily guess the right-side (test) molecular energy by looking at the left-side (training) ones. Of course, the children has not solved the Kohn-Sham equation; the children just look and guess. That is, this is just a pattern matching and I think this seems not to be physics at all.

As I have posted the previous question, I now understand that transferring to different properties is a very difficult task. However, I believe that a physically-meaningful AI/ML model should be able to be transferred to various physical properties; such a model is called "foundation model" in recent AI/ML. However, current research can not predict the hydrogen molecular energy if the training dataset does not include the hydrogen molecular energy.

Answer 1

"I am a student now and should decide my major to research soon. AI/ML is a very hot research field and I am very interested in it, but I have some doubts before I study AI/ML."

AI is indeed a hot topic, and there's plenty of reasons to have doubts before committing yourself 100% to a field, but:

"I found the GitHub issue of this paper, and the high-accurate ML model (called DimeNet, which is a graph neural network model) cannot predict the energy of hydrogen molecule at all (the error is over 100 kcal/mol!). The reason, according to the author of this paper, is that the training dataset (this paper used the benchmark QM9) does not include the hydrogen molecule."

But this is what the author actually said:

"From a physical perspective the H2 molecule is certainly the simplest molecule there is, but from the QM9 data perspective this is an extreme outlier. There are no H-H bonds in the QM9 dataset"

Saying that a dataset does not involve "any H-H bonds" is different from saying that it "does not include the hydrogen molecule".

You're making it seem like machine learning doesn't work unless it has been trained on exactly the same system for which you want to make a prediction, which is not true. For example, to predict the energy for C₂, the author is not claiming that you would need C₂ to be in the training set, as long as enough molecules with C-C bonds (like most organic molecules) were to be in the training set.

"I believe that even children can easily guess the right-side (test) molecular energy by looking at the left-side (training) ones."

Perhaps that's true in this one very specific case from a cherry-picked paper. It does not mean that all machine learning papers demonstrate no more than what "children" can predict on their own.

AI is a very hot field for a good reason. AI more than just machine learning. Machine learning is also more than just pattern recognition. For example reinforcement learning is what was used when AlphaGo beat a human at Go, and has applications in self-driving cars and other autonomous vehicles (the University of Waterloo already offers people on campus an autonomous bus).

In terms of matter modeling, DeepMind (the same company that made AlphaGo) made AlphaFold for protein folding, and AlphaFold has won competitions. DeepMind also made FermiNet which does seem to calculate variational electronic ab initio energies of some systems, which are lower than any other energies calculated (using for example variational Monte Carlo, or diffusion Monte Carlo, or direct full-configuration-interaction solvers). This is truly state-of-the-art research and its thanks to the enormous wealth of valuable insight that artificial intelligence researchers have brought to the world in recent years. Many other examples of applications can be found in our machine-learning tag.

Finally, I do very much appreciate that there's still reasons to be skeptical about how useful AI will be to matter modeling, but I would not at all discourage you from entering the field of AI. Jobs in AI are far more abundant than jobs in matter modeling and often pay 10x more (it is no exaggeration that a lot of AI researchers at Google, Apple, OpenAI and other places earn more than \$1 million/year). Even someone as young as Ian Goodfellow, who did not have a single paper published before 2009, had a salary of \$800,000/year almost immediately after completing his PhD.

Even outside of science, people working in law, education, writing, finance, and many other fields are increasingly using AI, and in many ways it's becoming a form of literacy. Saying that you're too skeptical about it would be like saying you're skeptical about learning math, or computer programming.

Answer 2

I did my master's research project in predicting absorption spectra using machine learning (with fingerprints). So, my research is somewhat in the field of material modelling you can say. I want to add my two cents to the discussion.

As shown in this figure, many data samples have very similar molecular structures and energies, so I believe that even children can easily guess the right-side (test) molecular energy by looking at the left-side (training) ones. Of course, the children has not solved the Kohn-Sham equation; the children just look and guess. That is, this is just a pattern matching and I think this seems not to be physics at all.

It is absolutely true that machine learning models are essentially looking for patterns in the training dataset. But I don't think it is a bad thing. Because looking for patterns is all we do, in fact, I would argue that it is the only thing we can do. You are making a distinction between physics and pattern matching, but I think this distinction is an illusion.

It's not like we have access to the source code of the universe. So what we do (our brains) is look for patterns of events that seem to follow the same rule. For example, we see that oppositely charged particles attract each other, and same charged particles repel. We look at the pattern, and we guess that there must be a rule the particles are following. Then we measure the forces, and find that we can write the rule down as a mathematical equation—this is the Coulomb force. But please notice that there has to be a pattern for us to perceive in the first place.

If that sounds too philosophical, look at it from a practical point of view. When we are dealing with forces between charged particles, the patterns are simple enough to be represented perfectly by a simple analytical expression. When we are predicting the energy of a molecule from its structure, there is no simple expresssion that we have found. But our intuition tells us there must be a pattern, so we use machine learning models. We feed the model huge amounts of data, in the hope that it can find the underlying pattern that is actually there. If the ML model does not work, that means it did not find the actual pattern, but it does not mean the whole methodology has to be discarded.

Ideally, your ML model would work on any new molecule that you throw at it, but in practice it never does. ML models work best if they have seen molecules similar to it in the training set (similar, not same). For example, in my research, I found that molecules had $< 0.02$ eV error in predicting excitation energies, if the training set had $> 10$ molecules with $>0.7$ similarity coefficient.

Answer 3

While machine learning is becoming very important, I think physics-based models continue to be necessary in matter modeling. ML models are good for systems where there is a lot of data available. On the other hand, physics-based models can often be applied for more exotic systems that have been poorly studied in the past. Indeed, physics-based models are often needed to supply data to the AI models.

For example, consider that you want the energy of the molecules in the study you cited, but instead of in gas phase, you want the energy at a silicon (111) surface. Or under a high magnetic field. Or instead of organic molecules you are interested in the energies of boron–oxygen and boron–nitrogen clusters. For these situations, you will probably find very little data available and will need to perform quantum chemistry calculations to get the data necessary to train the AI model.

So, if you like physical models, there is probably still plenty of room for developing expertise in physical models. However, I think there is no doubt you will also find ML to be useful in analyzing the results of the physical models and developing useful predictions from them.

But as ML and physics-based methods get increasingly complex, people will probably tend to specialize in one or other. In the future, matter modeling will probably have experts in ML who dabble in physics-based methods and experts in physics-based methods who dabble in ML.