Structural Analysis/Design (STRAND) and Neural Net Computation (NETCOM) are developmental modular computer programs that exploit the speedup afforded by parallel and neural-network computing to drastically reduce the computation time needed to solve large design-optimization problems. These programs were originally intended for use in designing aerospace structures, but, when fully developed, will also be useful for optimizing designs of diverse nonaerospace structures, including high-rise buildings, automobile structures, and civil infrastructures.

The PAR_STRAND_NET Software System utilizes a combination of neural-network, finite-element, stochastic-optimization, and parallel-computation techniques to reduce the computation time needed to solve large design-optimization problems.

STRAND and NETCOM have been integrated into PAR_STRAND_NET, which is a self-contained software system that can be implemented on either PVM or MPI standard networked parallel clusters of computer workstations and IBM SP computers. The figure shows the macro flow chart of PAR_STRAND_NET.

STRAND can perform automated design-optimization computations based on neural-network (NN), finite-element (FE), and mixed NN/FE analysis methods. STRAND affords computational capabilities to perform tasks as follows:

  • Multiple concurrent structural analyses can be performed by use of parallel multiple FE and/or NN analysis methods and modules.
  • Parallel optimization of structural design can be accomplished by use of FE and/or NN structural-analysis modules and an optimization module. The module implements the Integral Global /Local Optimization (IGLO) algorithm, which performs stochastic global and local searches. Built-in objective functions of structural weight, strain energy, and maximum displacement are used in optimization.
  • Training examples and data bases for training neural networks can be created by use of concurrent multiple FE structural-analysis modules and a scaled-training-example data-reduction module.

NETCOM is capable of training multiple NNs or sub-NNs in parallel and of predicting NN output quantities by use of trained NNs. Training in NETCOM is effected by the back-propagation algorithm. A capability for concurrent training of multiple NNs or sub-NNs can be utilized in cases in which (1) the NNs or sub-NNs share the same input vector (typically, the same set of cross-sectional areas of structural members or of other design variables) but (2) the NNs or sub-NNs generate different output vectors (e.g., displacement vs. stress vectors) or numerically differing components of the same output vector. Once NN training has been completed, NETCOM generates a second data base that contains the matrices of synaptic weights of the trained NNs.

The use of NNs to replace FE reanalysis in the optimization of a structural design can reduce computational time by nearly an order of magnitude — an important advantage in the case of a large-scale design. Ordinarily, this advantage would be offset by the tedious and time-consuming nature of NN training. However, in this system, computational burden of NN training is reduced by the use of reduced NN models plus the efficient parallel and NN computational capabilities of NETCOM.

On the basis of numerical performance in tests conducted thus far, the IGLO algorithm was found to be much more efficient than are such other stochastic algorithms as those of the genetic and simulated-annealing types.

Thus, it appears that IGLO offers great potential for solving large-scale design problems that involve not only continuous design variables but also discrete variables and mixes of continuous and discrete variables.

One likely goal of future development efforts would be to secure the advantages while avoiding the disadvantages of both gradient-based optimization (GBO) methods and stochastic methods like those of IGLO. GBO methods are inapplicable to mixed- and discrete-variable problems, and sometimes yield solutions that are not optimum in the sense that they correspond to local minima that differ from global minima of objective functions in design-variable space. On the other hand, whereas stochastic methods yield global solutions for continuous, discrete, and mixed variables, much more computation time is needed to find a local minimum in a stochastic method than in a GBO method. Therefore, it appears desirable to replace the continuous-local-search portion of the IGLO algorithm with a GBO algorithm to form an integrated IGLO/GBO algorithm to increase computational efficiency and the quality of solutions of problems that involve continuous, discrete, and mixed variables.

This work was done by Rong C. Shieh of MRJ Technology Solutions for Lewis Research Center. Inquiries concerning rights for the commercial use of this invention should be addressed to

NASA Lewis Research Center, Commercial Technology Office, Attn: Tech Brief Patent Status, Mail Stop 7-3, 21000 Brookpark Road, Cleveland, Ohio 44135

Refer to LEW-16601.