Tag: JGI Data Scientist Support

PGR JGI Seed Corn Funding Project Blog 2023/24: Binyu Cui

Introduction:

Magnetic components, such as inductors, play a crucial role in nearly all power electronics applications and are typically known to be the least efficient components, significantly affecting overall system performance and efficiency. Despite extensive research and analysis on the characteristics of magnetic components, a satisfactory first-principle model for their characterization remains elusive due to the nonlinear mechanisms and complex factors such as geometries and fabrication methods. My current research focuses on the characterization and modelling of magnetic core loss, which is essential for power electronics design. This research has practical applications in areas such as the fast charging of electric vehicles and the design of electric motors.

Traditional modelling methods have relied on empirical equations, such as the Steinmetz equation and the Jiles-Atherton hysteresis model, which require parameters to be curve-fitted in advance. Although these methods have been refined over generations (e.g., MSE and iGSE), they still face practical limitations. In contrast, data-driven techniques, such as machine learning with neural networks, have demonstrated advantages in addressing multivariable nonlinear regression problems.

Thanks to the funding and support from the JGI Institute, the interdisciplinary project “MagMap” has been initiated. This project encompasses testing platform modifications, database setup, and neural network development, advancing the characterization and modelling of magnetic core loss.

Outcome

Previously, a large-signal automated testing platform is produced to evaluate the magnetic characteristics under various conditions. Fig. 1 shows the layout of the hardware section of the testing platform and Fig. 2 shows the user interface of the software that is currently used for the testing. With the help of JGI, I have managed to update the automated procedure of the platform including the point-to-point testing workflow and the large signal inductance characterizing. This testing platform is crucial for generating the practical database for the further machine learning process as its automated function has largely increased the testing efficiency of each operating point (approx 6-8s per data point).

Utilizing the current database, a Long Short-Term Memory (LSTM) model has been developed to predict core loss directly from the input voltage. The model shows a better performance in deducing the core loss than traditional empirical models such as the improved generalized Steinmetz equation. A screenshot of the code outcome is shown in Fig. 3 and an example result of the model for one material is shown in Figure 4. A feedforward neural network has been tried out as a scalar-to-scalar model to deduce the core loss directly from a series of input scalars including the magnetic

flux density amplitude, frequency and duty cycle. Despite the accuracy of the training process, there are limitations in the input waveform types. Convolutional neural networks have also been tested before using the LSTM as a sequence-to-scalar model. However, the model size is significantly larger than the LSTM with hardly any improvement in accuracy.

Future Plan:

Although core loss measurement and modelling is a key issue in industrial applications, the reason behind these difficulties is the non-linear relationship between the magnetic flux density and the magnetic field strength which is also known as the permeability of the magnetic material. The permeability of ferromagnetic is very sensitive to a series of external parameters including temperature, induced current, frequency and input waveform types. With an accurate fitting between the relationship of magnetic flux density and field strength, not only

the core loss can be precisely calculated but also the current modelling method that is used in Ansys and COMSOL can be improved.

Acknowledgement:

I would like to extend my gratitude to JGI for funding this research and for their unwavering support throughout the project. I am also deeply thankful to Dr. Jun Wang for his continuous support. Additionally, I would also like to express my appreciation to Mr. Yuming Huo for his invaluable advice and assistance with the neural network coding process.

PGR JGI Seed Corn Funding Project Blog 2023/24: Zhijin Guo 

Introduction

The overall aim of this project is to analyse music scores by machine learning.  These of course are different from sound recordings of music, since they are symbolic representations of what musicians play.  But with encoded versions of these scores (in which the graphical symbols used by musicians are rendered as categorical data) we have the chance to turn these instructions in various sequences of pitches, harmonies, rhythms, and so on. 

What were the aims of the seed corn project? 

CRIM concerns a special genre of works from sixteenth century Europe in which a composer took some pre-existing piece and adapted the various melodies and harmonies in it to create a new but related composition. More specifically, the CRIM Project is concerned with polyphonic music, in which several independent lines are combined in contrapuntal combinations. As in the case of any given style of music, the patterns that composers create follow certain rules:  they write using stereotypical melodic and rhythmic patterns. And they combine these tunes (‘soggetti’, from the Italian word for ‘subject’ or ‘theme’) in stereotypical ways. So, we have the dimensions of melody (line), rhythm (time), and harmony (what we’d get if we slice through the music at each instant. 

We might thus ask the following kinds of questions about music: 

• Starting from a given composition, what would be its nearest neighbour, based on any given set of patterns we might chose to represent?  A machine would of course not know anything about the composer, genre, or borrowing involved in those pieces, but it would be revealing to compare what a machine might tell us about this such ‘neighbours’ in light of what a human might know about them. 

• What communities of pieces can we identify in a given corpus?  That is, if we attempt to classify of groups works in some way based on shared features, what kinds of communities emerge?  Are these communities related to Style? Genre? Composer? Borrowing? 

• In contrast, if we take the various kinds of soggetti (or other basic ‘words’) as our starting point, what can we learn about their context?  What soggetti happen before and after them?  At the same time as them?  What soggetti are most closely related to them? And through this what can we say about the ways each kind of pattern is used? 

Interval as Vectors (Music Ngrams) 

How can we model these soggetti?  Of course they are just sequences of pitches and durations.  But since musicians move these melodies around, it will not work simply to look for strings of pitches (since as listeners we can recognize that G-A-B sounds exactly the same as C-D-E).  What we need to instead is to model these as distances between notes.  Musicians call these ‘intervals’ and you could think of them like musical vectors. They have direction (up/down) and they have some length (X steps along the scale). 

Here is an example of how we can use our CRIM Intervals tools (a Python/Pandas library) to harvest this kind of information from XML encodings of our scores.  There is more to it than this, but the basic points are clear:  the distances in the score are translated into a series of distances in a table.  Each column represents the motions in one voice.  Each row represents successive time intervals in the piece (1.0 = one quarter note). 

Link Prediction 

We are interested in predicting unobserved or missing relations between pairs of ngrams in our musical graph. Given two ngrams (nodes in the graph), the goal is to ascertain the type and likelihood of a potential relationship (edge) between them, be it sequential, vertical, or based on thematic similarity. 

• Sequential is tuples that come near each other time.  This is Large Language Model which computes ‘context’. LLM then produces the semantic information that is latent in the data. 

• Vertical is tuples that happen at the same time.  It is ANOTHER kind of context. 

• Thematic is based on some measure of similarity.   

Upon training, the model’s performance is evaluated on a held-out test set, providing metrics such as precision, recall, and F1-score for each type of relationship. The model achieved a prediction accuracy of 78%. 

Beyond its predictive capabilities, the model also generates embeddings for each ngram. These embeddings, which are high-dimensional vectors encapsulating the essence of each ngram in the context of the entire graph, can serve as invaluable tools for further musical analysis.