For example, it is like this. In order to measure the

Recently, he participated as a defense judge for several doctoral candidates. They had something in common that they achieved research results by using deep learning, which has spread like wildfire in all fields of engineering in recent years, in their research. Interestingly, the candidates for defense took a very similar approach.

For example, it is like this. In order to measure the quantum field of a light-emitting material, an understanding of the ground state and the excited state of this material is required. When electrons in the ground state are excited by receiving external energy, electrons with high energy undergo a transition process to go back to the low energy state. In this transition process, electrons return from the high energy state to the low energy state, giving up as much energy as the difference. This energy can be absorbed by the phonon vibration of surrounding atoms, can be scattered to defects, or can be emitted neatly into photons without any obstacles. If the emitted light is in the visible light band, it can be seen even by our eyes. Light-emitting materials such as quantum dots, which are usually used in displays, are one of the main keys to how much energy is released back into light. Naturally, the more you put it out, the more desirable it will be.

What's interesting is that their excited states are not fixed as one. The probability distribution of excited states can be calculated using molecular orbitals for monomers, and DoS (density of states) calculated from the electron band structure for crystalline substances. Therefore, the luminous efficiency during the transition process will vary depending on the excited state. In any case, the lowest energy of the excited state acts as a kind of cliff edge, and the difference between the edge and the ground state becomes the photoluminescence we observe. Then, how can we separate the remaining energy that is not emitted in the form of photons? This can be seen by tracking and measuring whether these energies are reflected in the vibrations of various particles or quasiparticles that we know in very short periods of time. Of course, not all information can be thoroughly searched even with these experimental equipment. Also, even if energy is released in the form of photon emission, there is not only one method. Electrons can be emitted one by one, but electrons can be transferred in plural like one body like synchronized players, or electrons and holes can meet and fall. Each of them can be calculated by solving a time-dependent matrix dynamics equation using a transition matrix as a Hamiltonian in quantum mechanics. Here, the eigenvalues become parameters that represent these emission dynamics.

Therefore, in the experimentally measured PL spectra, it is assumed that several decays dynamics are involved in the transition from purely exciting state to ground state by initially separating the instrumental excitation. The neatest thing is that the excited state is simple and there is only one type of dynamics. Since they can be modeled in the form of I(t) ~I0*exp(-t/tau), which is a function dependent on the time of the intensity I of the PL, tau is calculated by calculating the slope A in the linear relationship between logI(t) vs At + B with a log function on both sides. (tau = -1/A), by the way, most light-emitting materials, especially samples involving organic semiconductors, may have multiple decays dynamics. Various quasi-particles such as small or large polaron, plasmon, and polariton, as well as basic cases such as electron, hole, multi-electron, and exciton, can participate in these decays dynamics. Since their dynamics time scales are different, in fact, if they are reflected in the experimental results, there will be several linearities or mixed in log-scale data. Therefore, care must be taken carefully enough when fitting time-resolved PL (TRPL) data.

In the defense, which he participated as a judge, the candidate showed the TRPL data of an organic material he synthesized and pointed this out well. So far, so good. However, he said that he did not know what physics might be hidden in this data, so he assumed it to be a black box and used physics-informed machine learning to estimate the various dynamics in it. So far, it was good. However, there were only three cases of the physics model he showed: 1 electron, 2 electrons, and exciton. In addition to this, several types of number density of defects were considered as additional bias. Using this 'physical model', the candidate made several sets of composite data using their weights, time-scale, delay, etc. as their main features. Based on this data set, he learned his MLP NN and completed a weight matrix composed of several layers, that is, the well-known convolutional structure NN. The candidate proudly analyzed the TRPL experiment data using this NN, and through this, single electrons, two electrons, excitons, and defect dynamics were separated and tabulated for several samples. My anxiety has crept in from here.

On the surface, it appears that any engineering graduate student did well in multivariate regression statistically by introducing machine learning to his research. Additionally, it appears to be flawless because it developed synthetic data based on its own 'physical' model and trained it, and it went through a test/validate process divided into 15:85 according to the standard. The issue is that the candidate believed that this was physics-informed machine learning.

After the candidate's presentation was over, I asked a few questions. The first question was why do you think it is physics-informed machine learning? The candidate was trained with data based on various models with an expression of naturalness, and since this model is well-known decoding dynamics, they responded that there was no problem. However, I asked again if it was really physics-informed and compared it with the decays dynamics time scale of single electron and multi-electron. He said he showed these figures in the table he organized and analyzed them sufficiently. However, the figures he showed in the table were only the figures shown by his NN, and it was not mentioned at all whether the figures made physical sense. I pointed out that both processes are electron-involved processes in common, but because the mechanisms are completely different, the order must be different and that there must be a dependence on the excitation power density. The candidate's expression seemed to start to color in embarrassment. It was because I thought it was a research result that both the supervisor and the person were satisfied with because it was already publicized.

There was another problem. The model considered by the candidate is not one that considers all physics. Of course, there is no model that 100% matches the truth because there are assumptions and simplifications at some stage. Still, it is necessary to consider all the mechanisms known so far. In addition to the charger caregiver, it is essential to consider these various processes in which other quasiparticles are involved, especially in organic matter, and the participation of these quasiparticles was completely blocked in the candidate's model. I asked the question. It is necessary to consider that in the figures produced by NN, for example, in this number filtered as only involved in single electrons, the long-range correlation effect of large polaron can be involved, especially in thin or small samples, and even more so in organic matter. The physics-informed physics that the candidate believed was actually just a toy model, but he seems to have believed that this was all the world in which the system he analyzed was involved.

In fact, not only candidate Lee but also quite a few students and researchers are still repeating this mistake. For example, if you only know the crystal structure, that is, the arrangement of atoms in a unit cell, you can find out the X-ray diffraction pattern by simulation. Mathematically, it is just a Fourier transform. However, experimentally obtained samples are usually not perfect single crystals. There are grain boundaries, there are many types of defects, and even very small samples (nano-scale), size and shape effects can be seen. In the case of a solid solution, the lattice constant shift may not follow only linearity, and symmetry is sometimes broken due to local strain. However, for this diffraction pattern, I have seen a case where I firmly believe that my sample is the perfect single crystal and learn the simulated pattern to estimate the composition of the solid solution as NN. This was also done because it was believed to be physics-informed NN, but in reality, it was not physics-informed, but approached & ideal mod