Comments about the article in Nature: The data-driven future of high energy density physics

Following is a discussion about this article in Nature Vol 593 20 May 2021, by Peter W. Hatfield e.a.
To study the full text select this link: https://www.nature.com/articles/s41586-021-03382-w

• The text in italics is copied from the article
  
• Immediate followed by some comments
  

In the last paragraph I explain my own opinion.

Contents

• 1. Chalenges
• 1.1 Experimentel design and automation
• 1.2. Data synthesis
• 1.3. physics Models
• 2. Case Studies
• 2.1 Astrophysics
• 2.2. Inertial Confinement Fusion (ICF)
• 2.3. Automation for high repitition rates
• 3. Policy propositions
• 3.1 Education
• 3.2. Research practices
• 3.3. Conclusion

Reflection

• Reflection 1 - AI and understanding.
• Reflection 2 -
• Reflection 3 -

Introduction

Who should be at the controls, however—humans, or artificial intelligences?

Always humans have the the overall responsibility to perform a certain task. In order to perform this task a mechanical and electronic equipment can be used, but the responsibilities stay with the designers i.e humans. If this electronic equipment are computers, executing software program does make no difference. If these programs are called AI, its all in the name.

In particular, phenomena at these conditions are highly nonlinear—small parameter changes can lead to large changes in behaviour.

Highly nonlinear systems are always difficult to control. Large changes in control parameters should be minimised in order to keep the system stable.

Interpreting extreme physics data typically requires simultaneously comprehending large amounts of complex multi-modal data from multiple different sources.

That is correct. The strategy should be to control the process in a step wise fashion i.e. not everything at once.

Optimizing extreme physics systems requires fine-tuning over large numbers of (often highly correlated) parameters.

Optimizing a complex physical system with many control parameters is very difficult. The most promising method is first to define a methematical model of the system. In that case the whole system can be simulated on a computer and the optimum soltion can be found. One question is what is it that you want to optimise. If

Key contemporary problems in HEDP theory include understanding multi-species plasmas, self-consistent emission, absorption and scattering of radiation, non-equilibrium plasmas, relativistic electron transport, magnetized plasmas, and quantum electrodynamic effects.

The primary strategy should be to study each first in 'isolation' and than in 'combination'.

As already discussed, understanding these conditions is key in astrophysics.

One more reason to follow this strategy

In this field, it is becoming possible to study particle and nuclear physics through HEDP experiments, as well as novel phenomena that are predicted to emerge only at extreme conditions, for example, the predicted thermal Schwinger process: the spontaneous production of electrons and positrons at very high electric field strengths.

Experiments are the primary source to understand all of physics. At the same time more detailed experiments are required to understand the production of electrons and positrons. There must be a reason why i.e. there must be a cause.

HEDP has a rich heritage of experimentation, currently practiced by thousands of scientists in several large facilities around the world.

That means a certain amount of coordination is required.

There are also important synergies with closely related areas of physics, for example, magnetic confinement fusion, solar probes and the detection of high-energy cosmic rays.

Exactly.

Some AI solutions from other fields are likely to be applicable in plasma physics—but HEDP also has its own unique challenges.

Okay.

Specifically, we typically want to (sometimes very rapidly) fine tune (either optimizing or fitting a model) a large number of parameters for a desired outcome, based on a large number of multi-modal datasets.

That depents about the time constraints within the process. You cannot boil a large amount of water instantaneous.

This is very difficult for humans, but even more difficult for a computer program.

This depents completely about the number of constraints embedded in the program i.e. all the rules of what to do and what not to do.

1. Chalenges

1.1 Experimentel design and automation

A key component of HEDP is experimentation.

That is the starting point for any endavour in science, if you want to understand the evolution of any process better.

Designing experiments is a hugely complex task, and researchers will typically have a wide range of overlapping goals during this phase.

This depents. Generally speaking, what we want to find out, is more detail in the evolution of the process studied. Suppose the process studied is a collision between two molecules. During the collision a sequence of reactions take place, producing individual particles in a sphere around the center of the point of collision. Suppose that only particles can be measured in a sphere at minimal 1 meter from the point of collision, that means intermediate events can not directly be measured. If you can measure the direction and speed of the different particles after 1 meter, still a certain 3D picture can be estimated about: what, where and when certain immediate results occured.
The problem is that in all measurements, all calculations and finally in all results uncertainties are involved

Of course the process to decide how to improve this particle experiment can de a hugh task, specific if more people want to use the equipment.

Researchers must consider what specific hypotheses are to be tested, and whether the expected data would be sufficient to rule out alternatives.

Very often when a certain apparatus is closed for maintenance, the 'only' task that the 'maintenance' team is trying to improve is accuracy.

In current design approaches there is typically much use of intuition, and often experiments are built as an extension of what has been done before, limiting the regions of experimental space that are studied.

I doubt if intuition is the right driving force to do science. Of course when the final experiment has done the whole team has to come to a certain conclusion, by taking all the facts into account what to do next. As part of that process a document must be generated about: What have we finally received, what went wrong, what actions for improvement are possible etc.
Next:

Machine learning techniques offer a possible framework within which the intuition of the computational scientists and experimental scientists can be explicitly included in a cohesive picture that considers both measurements that can be made, and which aspects of physics have the most leverage on those measurements.

The main task what to do next is a function of the physicists and involves physical issues. To what extend a computer program can be used in anyway to help to decide, from a physical point of view, what to do or give advice I have my doubts.
To do the same but now related to calculations involving measurements I also have my doubts.

AI-aided design is beginning to be used in the creation of new HEDP experiments, and we foresee that becoming the norm in the coming years.

TO assume that computer programs can be used in designing experiments I also have my doubts.

However, machine learning methods have yet to demonstrate that they can ‘think outside the box’ of pre-defined parameter spaces, so for the foreseeable future it will still be necessary to have substantial human input into the design process.

THAT SAYS IT ALL. I expect human input will always be required!

Experiments on state-of-the-art high-repetition-rate lasers, firing many times a second, cannot be done with a human in the loop, so in this case some algorithmic control is essential.

More detail about these types of experiments should be explained. In the case of a fighting machine there is a pilot which tries to control the airplane in such a way that the rear gunner can hit the plane of the enemy. The task of the rear gunner was to return the fire immediatly with a long burst. (See: Another Gunners Tale 10/11 April 1944 I D Hancocks (Taffy): https://tailendcharlietedchurch.wordpress.com/tail-end-charlie-ted-church/rear-gunner-sentinel-of-the-skies/
The point is to control such a burst of an automated machine gun or a laser no human is in that loop. See https://en.wikipedia.org/wiki/Machine_gun#Design That means humans can only be involved is outside that loop i.e. to select the next target.

1.2. Data synthesis

To isolate individual aspects of the underlying physics, experiments therefore typically require multiple, indirect observations, sometimes spread across several different experimental facilities.

That are the standard problems. That is all in the game. The more this can be automated the better, but has nothing to do with AI.

At most facilities researchers have developed multiple diagnostics for experiments; for example, both X-ray and particle spectra may be measured on a single experiment, along with many other forms of experimental data, all of which might contribute to the determination of a single quantity.

Again this is normal. If a single quantity is calculated the mathematics involved can be very important, because certain important factors can be singled out.

The analysis of such increasingly sophisticated interlinked data requires the use of more advanced modelling techniques to make the best use of the available data, and to quantify the uncertainty on any inferences in a sensible manner.

The problem with any complex interlinked system is to fully understand how the in principle independent subsystems interfer which each other. The problem of uncertainty is not a direct issue by taking care for the necessary forms of feedback. That is the case for systems with only humans as for systems which are centered around individual PC's or automation centers. See for example Fig 4.

1.3. physics Models

Owing to numerical approximations, poorly known or unknown model parameters, and missing physics (model discrepancy), computer models often do not accurately represent the physical process under study.

That is normal. Mathematical models i.e. mathematical equations are always an approximation of the physical process or the physical reality.

We can leverage real-world experiments to calibrate our computational models, enabling us to constrain some of the uncertain model parameters; ideally including an uncertainty quantification analysis and practising ‘data assimilation that obeys physical laws.

Physical laws are also mathematical equations. The parameters of these equations, in many cases, can be calibrated by experiments and observations. The accuracy of these observations is reflected in the uncertainty of the calculated parameters.

2. Case Studies

2.1 Astrophysics

With these vast quantities of observational and experimental data, the natural next step is to use experimentally calibrated models of microphysics like these in astrophysical models—with the potential to give improved predictions over theory-only models.

IMO any state of the art astrophysical model is a model (set of mathematical equations) that is based on the evolution in time of our understanding of the astro-physics incorporated in the interior of stars or blackholes. The result of present day observations and experiments, is to improve this astrophysical model. To do that is difficult.
What this means that the name theory-only models is wrong. Both are mathematical equations. See also: Reflection 1 - AI and understanding.

There is a huge amount of data on astrophysical plasmas of a wide variety of sources, taken in a huge variety of ways, but unfortunately they are currently held in different forms, by different communities.

Okay

See Fig. 2 for an infogram on how different astrophysical datasets could in future be combined in a data assimilation framework

This means in the future. How to do that in real is not easy.

Fig. 2: Integration of astrophysical information. At the center of this system is an "An AI predictive model" or an "Machine learning algorithm that is trained on data from all possible sources and is restricted to obey physical laws" Why should it obey physical laws? The problem can be in the existing physical laws. Input comes from: Solar space missions Telescopes. Observational data on the full range of astrophysical bodies. Laboratory astrophysics. Output goes to: Improved astrophysical inference. Better inferred astrophysical properties. Improved space weather forecasting Improved experimental predictions. More acurate experiments. Better control over conditions.

2.2. Inertial Confinement Fusion (ICF)

ICF experiments present distinct data challenges owing to their scale and complexity.

Okay

Experiments are also highly integrated, meaning that direct measurement of any figure of merit beyond the raw energy yield is not possible.

That is a honest remark, however what does that mean in practice.

These factors mean that ICF datasets are always sparse, with multiple confounding factors and uncertain information content; researchers therefore place a very high value on theoretical studies undertaken using multiphysics simulation codes.

The first half shows that the quality of the ICF observations are sparse and of low quality, when that is the case to use only theoretical studies to make progress in physics, is not the path to follow.
The line of 'attack' to follow is to improve the quality of observations.

Although cheaper than experiments, the simulations are still expensive, requiring at least months of central processing unit time to complete.

Simulations based on unreliable data means throwing one's money down the drain. The first strategy is to improve the experiments and try to interpret the results with a pencil and a piece of paper.

"par 2.2.1"

There is a great need for methods that can help with experimental design and optimization, interpretation of experimental data, linking experiments with physics models, as well as making reliable predictions of future experiments.

I think what people want are better managers and better scientist, with a smile, and when you have hired all of them you still are not satisfied, with the chance of getting a burn out.
It is very difficult to write a computer program to link the results of an experiment with one physical model or models in general.
See also: Reflection 1 - AI and understanding.

As this work progresses, ICF is becoming a prototypical example of the difficulties associated with science in the data-poor regime.

Reading the above text, people want to much.

As with the other examples in this Perspective, the fundamental data problem is the synthesis of multiple sources of information.

Okay

Here, the key aim is to efficiently use the sparse information available to update our physics understanding in order to make simulations more predictive of future experiments.

2.3. Automation for high repitition rates

To succeed, the automation of experiments requires both control of experimental parameters and real-time analysis of experimental results in one single algorithmic process. see Fig. 4.

An algoritmic process does not exist.

3. Policy propositions

3.1 Education

3.2. Research practices

3.3. Conclusion

Reflection 1 - AI and understanding.

When you read the discussed article, which is extremely valuable, you are left with the question, what is written, does that describe the present situation or is it more what people want? And if the article describes the future: Is it realistic? I have my doubts.

In paragraaf 2.3.1 the relation between AI and understanding is discussed.
In paragraaf 2.3.2 the role of researchers is discussed.
The major problem is that programs don't have the capability to understand something. You can call the task that a program performs: AI, but it is still a program. Only humans have the capability to understand. What a program does it executes a program, for example it can calculate a set of equations which are used to simulate a physical process, for example an explosion or firework. Once the program is written, the only thing that program can do is to simulate the physical process i.e. an explosion.
Normally the equations have parameters. The first task of the simulation is to calculate the parameters such that the results of the simulation match the results (observations) of an actual explosion. In a case of a firework the whole process consists of many explosions in 3D space. To simulate a firework the simple program that simulates one explosion is not enough; it has to be modified and adapted. That is the task of the researchers. In general that is situation with all programs; sooner or later you come to the limits of what the program can do and it requires modifications i.e. extensions.
Of course before your original program that simulates an explosion is declared completed you can also already implement certain extension in your program. That is possible. But that means that the intelligence of your program is exactly as intelligent as the researchers involved.
However there exists a snake in the garden. Suppose you want to simulate a process that is in conflict with the standard model using some 'new' type of physics. IMO that is impossible, because this requires some form of new particles which require new mathematical equations including new parameters. IMO you have to be clairvoyant. What is worse it is very difficult to test such a program, because we have no glue what the observations are of the processes in which these new particles are involved.
The bottom line is, that at the present it is impossible to write programs which are more intelligent (describe the future) than those which are based on the present day experiments.

Reflection 2

Back to my home page Index
Back to Nature comments Nature Index