SGML 2.0.0

A package for solution-guided machine learning method

---

Welcome to use SGML!

---

1 Introduction

Solution-guided machine learning (SGML) is a universal approach designed to enhance the extrapolation capabilities of AI models, as detailed in the paper [Thin-Walled Struct. 200 (2024) 111984]. To simplify the integration of this method into diverse projects, we have encapsulated it within a user-friendly Python package. By leveraging the functions provided in the package, users can effortlessly apply the SGML method to enhance the extrapolation capabilities of their AI models.

The current version of the package incorporates various regression models, including implementations for artificial neural networks, support vector regression, AdaBoost regression, Bayesian Ridge regression, and Ridge regression models. It's worth noting that we are actively working to expand the package's capabilities, and future releases will introduce additional models to further enhance its universality and applicability to a broader range of AI projects.

2 Preparation

• The functionality of this package depends on the following external libraries:

                                            

• You can easily install SGML using the following command (soon):

  pip install SGML

[!IMPORTANT] Currently, you can decompress the "SGML_Package.zip" file located in the appendix and then copy the decompressed file into your Python library to utilize it.

• The required project data is in .csv format, displayed in the table below, with the column index encompassing the features and labels necessary for machine learning.

  ID	x1	x2	...	y
  1	...	...	...	...
  2	...	...	...	...
  ...	...	...	...	...

3 API

[!IMPORTANT]

• The solution function will be returned with the data type of a function for the given solution. Users can assign it a name for subsequent guidance in machine learning.

• All models are presently implemented as functions, each returning the trained model along with both the labels and predictions for the testing set.

• When the parameters are set to 'default' or left unspecified, the default values for those parameters will be applied.

3.1 Solution Function

def SGML.create_solution_function(

expression = str,

variables = list

)

return function

[!TIP] expression : Solution expression, such as 'a**3+2*b+1'.

variables : List of variables included in the solution锛宻uch as ['a', 'b'].

3.2 Artificial Neural Network-based Model

def SGML.ann(

train_path = str,

test_path = str,

feature_names = list,

lable_names = list,

solution_functions = list,

model_loadpath = str,

model_savepath = str,

hidden_layers = list,

activation_function = object,

batch_size = int,

criterion = object,

optimizer = object,

learning_rate = float,

epochs = int

)

return object, ndarray, ndarray

[!TIP] train_path : The file path for loading the training set.

test_path : The file path for loading the testing set.

feature_names : List containing feature names, such as ['x1', 'x2', ...].

lable_names : List containing label names, such as ['y'].

solution_functions : List containing solution function names, such as [solution1, solution2, ...]. default=None

model_loadpath : The file path for the existing model. default=None

model_savepath : Path to save the model. default=None

hidden_layers : The hidden layer architecture, denoted as [4, 8, 2], signifies the presence of three hidden layers with node counts of 4, 8, and 2, respectively. default=[8, 8]

activation_function : The activation function鈥攔efer to PyTorch Documentation for details. default=torch.nn.PReLU()

batch_size : The number of training samples used by the model during each parameter update. default=Total number of samples

criterion : The loss function鈥攔efer to PyTorch Documentation for details. default=torch.nn.MSELoss()

optimizer : The optimizer鈥攔efer to PyTorch Documentation for details.default=torch.optim.Adam()

learning_rate : default=0.01

epochs : default=5000

3.3 Support Vector Regression-based Model

def SGML.svr(

train_path = str,

test_path = str,

feature_names = list,

lable_names = list,

solution_functions = list,

model_loadpath = str,

model_savepath = str,

kernel = str,

degree = int,

gamma = str or float,

coef0 = float,

tol = float,

C = float,

epsilon = float,

shrinking = bool,

cache_size = float,

verbose = bool,

max_iter = int

)

return object, ndarray, ndarray

[!TIP] The API reference for the parameters train_path, test_path, feature_names, lable_names, solution_functions, model_loadpath, and model_savepath can be found in Section 3.2.

kernel : Refer to sklearn.svm.SVR for detailed information, and the same applies to the following parameters. default='linear'

degree : default=3

gamma : default='scale'

coef0 : default=0.0

tol : default=1e-3

C : default=1.0

epsilon : default=0.1

shrinking : default=True

cache_size : default=200

verbose : default=False

max_iter : default=-1

3.4 AdaBoost Regressor-based Model

def SGML.adaboost(

train_path = str,

test_path = str,

feature_names = list,

lable_names = list,

solution_functions = list,

model_loadpath = str,

model_savepath = str,

estimator = object,

n_estimators = int,

learning_rate = float,

loss = str,

random_state = int

)

return object, ndarray, ndarray

[!TIP] The API reference for the parameters train_path, test_path, feature_names, lable_names, solution_functions, model_loadpath, and model_savepath can be found in Section 3.2.

estimator : Refer to sklearn.ensemble.AdaBoostRegressor for detailed information, and the same applies to the following parameters. default=LinearRegression()

n_estimators : default=50

learning_rate : default=1.0

loss : default='linear'

random_state : default=None

3.5 BayesianRidge Regressor-based Model

def SGML.bayesianridge(

train_path = str,

test_path = str,

feature_names = list,

lable_names = list,

solution_functions = list,

model_loadpath = str,

model_savepath = str,

max_iter = int,

tol = float,

alpha_1 = float,

alpha_2 = float,

lambda_1 = float,

lambda_2 = float,

alpha_init = float,

lambda_init = float,

compute_score = bool,

fit_intercept = bool,

copy_X = bool,

verbose = bool

)

return object, ndarray, ndarray

[!TIP] The API reference for the parameters train_path, test_path, feature_names, lable_names, solution_functions, model_loadpath, and model_savepath can be found in Section 3.2.

max_iter : Refer to sklearn.linear_model.BayesianRidge for detailed information, and the same applies to the following parameters. default=None

tol : default=1e-3

alpha_1 : default=1e-6

alpha_2 : default=1e-6

lambda_1 : default=1e-6

lambda_2 : default=1e-6

alpha_init : default=None

lambda_init : default=None

compute_score : default=False

fit_intercept : default=True

copy_X : default=True

verbose : default=False

3.6 Ridge Regressor-based Model

def SGML.ridge(

train_path = str,

test_path = str,

feature_names = list,

lable_names = list,

solution_functions = list,

model_loadpath = str,

model_savepath = str,

alpha = float,

fit_intercept = bool,

copy_X = str or bool,

max_iter = int,

tol = float,

solver = str,

positive = bool,

random_state = int

)

return object, ndarray, ndarray

[!TIP] The API reference for the parameters train_path, test_path, feature_names, lable_names, solution_functions, model_loadpath, and model_savepath can be found in Section 3.2.

alpha : Refer to sklearn.linear_model.Ridge for detailed information, and the same applies to the following parameters. default=1.0

fit_intercept : default=True

copy_X : default=True

max_iter : default=None

tol : default=1e-4

solver : default='auto'

positive : default=False

random_state : default=None

4 Example

As depicted in the figure below, considering that the fixed end of a cantilever beam is obscured, force-deflection data can be experimentally obtained from the visible end. Our objective is to leverage machine learning to glean insights from this data and predict deflection at higher forces in different locations. Interestingly, by treating the occluded part as a wall, we can readily calculate the deflection of the new cantilever beam, providing valuable guidance for machine learning applications.

Assuming $l=0.3 m$, $E=210 GPa$, $I=180 m^{-4}$. We will proceed to address this problem using the SGML.ann() module, and the corresponding code and results are presented below:

import SGML

# Define the solution function

solution_1 = SGML.create_solution_function(expression='F * (-1 / 6 * x ** 3 + 1 / 2 * 0.3 * x ** 2) / (2100 * 180)',
                                           variables=['F', 'x'])


# Call the SGML.ann() module for training

my_ann, y_test, y_pre = SGML.ann(train_path='./bending_train1.csv',
                                 test_path='./bending_test1.csv',
                                 feature_names=['F', 'x'],
                                 lable_names=['y'],
                                 solution_functions=[solution_1],
                                 model_loadpath='default',
                                 model_savepath='default',
                                 hidden_layers='default',
                                 activation_function='default',
                                 batch_size='default',
                                 criterion='default',
                                 optimizer='default',
                                 learning_rate='default',
                                 epochs='default')

The required data for this example is readily available within the package. For comparison, we set solution_functions=None (indicating conventional machine learning), retrain, and present the prediction results below.