Method and apparatus for solving complex and computationally intensive inverse problems in real-time

• US 6,208,982 B1
• Filed: 07/30/1997
• Issued: 03/27/2001
• Est. Priority Date: 11/18/1996
• Status: Expired due to Fees

• Current Assignee: Lockheed Martin Energy Research Corporation (Martin Marietta Corporation)
• Original Assignee: Lockheed Martin Energy Research Corporation (Martin Marietta Corporation)
• Inventors: Ivezic, Nenad, Zacharia, Thomas, Allen, John D. Jr., Ludtka, Gerard M.
• Primary Examiner(s): Hafiz, Tariq R.
• Assistant Examiner(s): STARKS, WILBERT L
• Application Number:
  US08/903,068
• Time in Patent Office: 1,336 Days
• Field of Search: 706/1, 706/11, 706/14, 706/15, 706/21, 706/23
• US Class Current: 706/11
• CPC Class Codes: G06F 30/00 Computer-aided design [CAD]

Abstract

The system of the present invention may “solve” a variety of inverse physical problem types by using neural network techniques. In operation, the present invention may generate data sets characterizing a particular starting condition of a physical process (such as data sets characterizing the parameters of an initial metal die), based upon an ending condition of the physical process (such as the parameters of the metal part to be stamped by the die). In one embodiment, the system of the present invention may generate a plurality of training data sets, each training data set characterizing a sample ending condition, the physical process that results in the sample ending condition, and a sample starting condition of the physical process. The training data sets may then be applied to a neural network so as to train the network. A network definition associated with the trained neural network may be stored, and an ending data set characterizing an ending condition of the physical process may be generated. A starting data set characterizing a starting condition of the physical process may thereafter be generated based upon the stored network definition and the ending data set.

Claims

1. A system for generating parameters characterizing a first physical object, wherein the first physical object is adapted according to a specified transformation process so as to modify a second physical object, the system comprising:

(a) means for generating a plurality of sets of training parameters, each set of training parameters characterizing;

(i) a sample second physical object,

(ii) the specified transformation process that results in the sample second physical object, and

(iii) a sample first physical object that interacts with the specified transformation process so as to modify the sample second physical object;

(b) means for operating a neural network;

(c) means for applying the plurality of sets of training parameters to the neural network so as to train the neural network, wherein each of the plurality of sets of training parameters is processed by the applying means over a plurality of patches, each patch comprising a data array, two orthogonal axes crossing the data array and a plurality of regions along the axes, so as to provide at least one single output node by averaging over the plurality of regions without losing information contained in the data array, thereby producing a network definition associated with the trained neural network;

(d) means for storing the network definition;

(e) means for generating a set of second parameters characterizing the second physical object; and

(f) processing means for generating a set of first parameters characterizing the first physical object based upon the set of second parameters and the network definition stored by the storing means.

2. The system of claim 1, further comprising means for creating the first physical object based upon the set of first parameters.

3. The system of claim 1, wherein the first physical object comprises a metal die, and the second physical object comprises a metal part stamped by the metal die.

4. The system of claim 1, wherein the neural network comprises a backpropagation network having at least one input processing element and at least one output processing element.

5. The system of claim 1, wherein the means for operating the neural network comprises a supercomputer.

6. The system of claim 1, wherein the processing means comprises a desk top computer.

7. The system of claim 1, wherein the training parameter generating means generates the plurality of sets of input parameters by modeling the specified transformation process using at least one method selected from the group consisting of Finite Element Analysis, Discrete Event Simulations, and Massively Parallel Processing.

8. A method for creating a first physical object, wherein the first physical object is adapted according to a specified transformation process so as to modify a second physical object, the method comprising the steps of:

(a) generating a plurality of sets of training parameters, each set of training parameters characterizing;

(i) a sample second physical object,

(ii) the specified transformation process that results in the sample second physical object, and

(iii) a sample first physical object that interacts with the specified transformation process so as to modify the sample second physical object;

(b) applying the plurality of sets of training parameters to a neural network so as to train the neural network, wherein each of the plurality of sets of training parameters is processed by a supercomputer over a plurality of patches, each patch comprising a data array, two orthogonal axes crossing the data array and a plurality of regions along the axes, so as to provide at least one single output node by averaging over the plurality of regions without losing information contained in the data array, thereby producing a network definition associated with the trained neural network;

(c) storing in a memory the network definition associated with the trained neural network;

(d) generating a set of second parameters characterizing the second physical object;

(e) generating at a desktop computer a set of first parameters characterizing the first physical object based upon the set of second parameters and the network definition stored by the memory; and

(f) creating the first physical object based upon the set of first parameters.

9. The method of claim 8, wherein the first physical object comprises a metal die and the second physical object comprises a metal part.