Machine learning approach to piecewise linear interface construction
Methods, systems, and apparatus, including medium-encoded computer program products, for a machine learning approach for piecewise linear interface construction include: obtaining a cell fraction for a mesh cell in a mesh, a normal vector perpendicular to a linear interface for the mesh cell, and geometry information of the mesh cell, wherein the geometry information includes at least two data values; normalizing the geometry information of the mesh cell to reduce the at least two data values to at least one data value; obtaining an output value from a machine learning algorithm, previously trained using normalized geometry information of multiple different cells of a same mesh type as the mesh cell; determining the linear interface for the mesh cell based on the output value, the normal vector and the geometry information; using the linear interface for the mesh cell in numerical simulation processing of the mesh.
1 . A method comprising:
obtaining, in a computer program, properties of interface cells from a set of mesh cells in a computer model;
providing, by the computer program, the properties of the interface cells to a machine learning algorithm, wherein the properties of the interface cells comprise a normal vector and a cell fraction for each of the interface cells, wherein the machine learning algorithm was previously trained using multiple different cells of a same mesh type as the interface cells from the set of mesh cells;
determining, by the computer program, a linear interface, which divides the interface cells, using output values from the machine learning algorithm; and
using, by the computer program, the linear interface during processing of the computer model.
2 . The method of claim 1 , wherein the same mesh type is a regular mesh cell type, and the properties of the interface cells consist of the normal vector and the cell fraction for each of the interface cells.
3 . The method of claim 1 , wherein the same mesh type is an irregular mesh cell type, and the properties of the interface cells comprise at least one data value representing normalized geometry information for each of the interface cells.
4 . The method of claim 3 , comprising: normalizing, by the computer program, initial geometry information for each of the interface cells to reduce a number of data values representing the normalized geometry information in the properties of the interface cells.
5 . The method of claim 4 , wherein each of the interface cells has three or more vertices, and
wherein the normalizing comprises: mapping the initial geometry information of each interface cell to a new coordinate system; and using coordinate information of one or more vertices of the normalized geometry information for the interface cell in the new coordinate system, wherein a number of the one or more vertices in the new coordinate system is less than a number of the three or more vertices; and
wherein the determining comprises: obtaining final output values by remapping the output values back to a coordinate system for the interface cells; and determining the linear interface using the final output values.
6 . The method of claim 5 , wherein the same mesh type is triangular, each interface cell is a triangular interface cell having three vertices, and the one or more vertices of the normalized geometry information for the interface cell is one vertex.
7 . The method of claim 1 , further comprising training the machine learning algorithm, wherein training the machine learning algorithm comprises:
generating multiple different cells of the same mesh type as the interface cells as training inputs;
obtaining, for each of the training inputs, a respective target output for the training input; and
training the machine learning algorithm on the training inputs and the corresponding target outputs.
8 . The method of claim 1 , wherein using the linear interface during the processing of the computer model comprises using the linear interface for the interface cells in numerical simulation processing of the computer model for computational fluid dynamics simulation, plastic injection molding simulation, or topology or geometry optimization.
9 . The method of claim 1 , wherein using the linear interface during the processing of the computer model comprises using the linear interface for the interface cells to extract a polyhedral mesh surface from a three-dimensional iso-surface defined in the computer model for use in mesh-based calculation or visualization.
10 . A system comprising:
a first computer communicatively coupled with a network, the first computer comprising a first non-transitory storage medium having instructions of a computer program stored thereon; and
a second computer communicatively coupled with the network, the second computer comprising a second non-transitory storage medium having instructions of the computer program stored thereon;
wherein the computer program is configured to cause the first and second computers to
obtain properties of interface cells from a set of mesh cells in a computer model,
provide the properties of the interface cells to a machine learning algorithm, wherein the properties of the interface cells comprise a normal vector and a cell fraction for each of the interface cells, wherein the machine learning algorithm was previously trained using multiple different cells of a same mesh type as the interface cells from the set of mesh cells,
determine a linear interface, which divides the interface cells, using output values from the machine learning algorithm, and
use the linear interface during processing of the computer model.
11 . The system of claim 10 , wherein the same mesh type is a regular mesh cell type, and the properties of the interface cells consist of the normal vector and the cell fraction for each of the interface cells.
12 . The system of claim 10 , wherein the same mesh type is an irregular mesh cell type, and the properties of the interface cells comprise at least one data value representing normalized geometry information for each of the interface cells.
13 . The system of claim 12 , wherein the computer program is configured to cause the first and second computers to normalize initial geometry information for each of the interface cells to reduce a number of data values representing the normalized geometry information in the properties of the interface cells.
14 . The system of claim 13 , wherein each of the interface cells has three or more vertices, and
wherein the computer program is configured to cause the first and second computers to
map the initial geometry information of each interface cell to a new coordinate system,
use coordinate information of one or more vertices of the normalized geometry information for the interface cell in the new coordinate system, wherein a number of the one or more vertices in the new coordinate system is less than a number of the three or more vertices,
obtain final output values by remapping the output values back to a coordinate system for the interface cells, and
determine the linear interface using the final output values.
15 . The system of claim 14 , wherein the same mesh type is triangular, each interface cell is a triangular interface cell having three vertices, and the one or more vertices of the normalized geometry information for the interface cell is one vertex.
16 . The system of claim 10 , wherein the computer program or another computer program is configured to cause the first and second computers to train the machine learning algorithm, comprising:
generate multiple different cells of the same mesh type as the interface cells as training inputs,
obtain, for each of the training inputs, a respective target output for the training input, and
train the machine learning algorithm on the training inputs and the corresponding target outputs.
17 . The system of claim 10 , wherein the computer program is configured to cause the first and second computers to use the linear interface for the interface cells in numerical simulation processing of the computer model for computational fluid dynamics simulation, plastic injection molding simulation, or topology or geometry optimization.
18 . The system of claim 10 , wherein the computer program is configured to cause the first and second computers to use the linear interface for the interface cells to extract a polyhedral mesh surface from a three-dimensional iso-surface defined in the computer model for use in mesh-based calculation or visualization.
19 . A non-transitory computer-readable medium encoding instructions operable to cause data processing apparatus to perform operations comprising:
obtaining, in a computer program, properties of interface cells from a set of mesh cells in a computer model;
providing, by the computer program, the properties of the interface cells to a machine learning algorithm, wherein the properties of the interface cells comprise a normal vector and a cell fraction for each of the interface cells, wherein the machine learning algorithm was previously trained using multiple different cells of a same mesh type as the interface cells from the set of mesh cells;
determining, by the computer program, a linear interface, which divides the interface cells, using output values from the machine learning algorithm; and
using, by the computer program, the linear interface during processing of the computer model.
20 . The non-transitory computer-readable medium of claim 19 , wherein the same mesh type is an irregular mesh cell type, and the properties of the interface cells comprise at least one data value representing normalized geometry information for each of the interface cells.
21 . The non-transitory computer-readable medium of claim 20 , the operations comprising: normalizing, by the computer program, initial geometry information for each of the interface cells to reduce a number of data values representing the normalized geometry information in the properties of the interface cells.
22 . The non-transitory computer-readable medium of claim 21 , wherein each of the interface cells has three or more vertices, and
wherein the normalizing comprises: mapping the initial geometry information of each interface cell to a new coordinate system; and using coordinate information of one or more vertices of the normalized geometry information for the interface cell in the new coordinate system, wherein a number of the one or more vertices in the new coordinate system is less than a number of the three or more vertices; and
wherein the determining comprises: obtaining final output values by remapping the output values back to a coordinate system for the interface cells; and determining the linear interface using the final output values.
23 . The non-transitory computer-readable medium of claim 22 , wherein the same mesh type is triangular, each interface cell is a triangular interface cell having three vertices, and the one or more vertices of the normalized geometry information for the interface cell is one vertex.
24 . The non-transitory computer-readable medium of claim 19 , wherein the same mesh type is a regular mesh cell type, and the properties of the interface cells consist of the normal vector and the cell fraction for each of the interface cells.
25 . The non-transitory computer-readable medium of claim 19 , the operations further comprising training the machine learning algorithm, wherein training the machine learning algorithm comprises:
generating multiple different cells of the same mesh type as the interface cells as training inputs;
obtaining, for each of the training inputs, a respective target output for the training input; and
training the machine learning algorithm on the training inputs and the corresponding target outputs.
26 . The non-transitory computer-readable medium of claim 19 , wherein using the linear interface during the processing of the computer model comprises using the linear interface for the interface cells in numerical simulation processing of the computer model for computational fluid dynamics simulation, plastic injection molding simulation, or topology or geometry optimization.
27 . The non-transitory computer-readable medium of claim 19 , wherein using the linear interface during the processing of the computer model comprises using the linear interface for the interface cells to extract a polyhedral mesh surface from a three-dimensional iso-surface defined in the computer model for use in mesh-based calculation or visualization.