(548b) Computational Optimization of Difficult Physics-Based Project Designs Using Massively Parallel Solution Methods

Environmental remediation has been ongoing in the United States for several decades and has produced some successes. However, the more challenging sites remain resistant to efficient and effective remediation. These are precisely the environmental problems that remain as persistent fiscal and technological challenges as the success of the heuristic trial-and-error approach with simplified conceptual site models, which assisted in remediating some of the simpler sites, has far over-reached its limit of applicability. The tools herein enable formal, physically consistent and robust analysis of optimal remedial designs for environmental contamination problems of any level of complexity including formal evaluation of uncertainty in design performance.

This paper discusses and demonstrates a massively parallel, and efficient computational optimization technology used for optimally managing environmental remediation project and programs by reducing remediation design analysis timeframes and costs. Those responsible for managing or remediating chemical or radionuclide impacted sediments, soils, surface or groundwater including designing remedial design and operation projects benefit from tools that help them find acceptable, efficient and effective solutions to complex and multi-faceted planning challenges at efficient cost points. HydroGeoLogic, Inc. (HGL) has extended and unified its universal optimization algorithm deployed in the Physics-based Management Optimization PBMO^TM system (Deschaine, 2016). This new optimization algorithm is successfully deployed on environmental and radionuclide impacted sites across a variety of computational environments including the high performance elastic Cloud-based computing platforms.

This computational advancement removes the impediment of burdensome and unpredictable runtime requirements for optimizing the design of these types of projects, which have typically consumed weeks or months of calendar time using heuristic trial-and-error or even automated serial computation approaches. These new optimization methods return results in minutes to hours for simulations and optimal solutions in hours to at most a few days for medium and larger size projects which here-to-fore required months or more to solve. This work recently received the 2017 Grand Prize Award (Research category) from the American Academy of Environmental Engineers and Scientists (AAEES, 2017a).

Computational Approach

PBMO^TM now includes a massively parallel single universal global optimization algorithm. Capabilities include optimizing linear or non-linear elliptic, parabolic and hyperbolic models and objective functions that are a continuous, non-continuous and mixed integer. It is a tool system that helps users perform complex high-level tasks with relatively amazing simplicity. Modes of operation include standard and expert modes, and it runs on computers with single and multiple CPUs, including local area grids and the Cloud. Because it is programmed as a set of modules using standard FORTRAN, Python and web languages, it is platform independent. It has the capability for including applications with mixed operating systems and can deploy with multiple, independent or blended process models including subject matter expert (SME), data-driven (DD), Physics-based (PB) or integrated model (IM). Analysis can begin with the best result from a trial-and-error approach, or from a random starting point, so efforts engaged to date by the design teams are fully used. The optimal solution search relies on evaluations of candidate solutions, which can be entirely physics-based or response function or minorant approximated solution search (Deschaine, 2003, 2013 and 2014).

Optimization is accomplished by implementing HGL's massively parallel optimal search strategy. The parallel distributed optimization method now regularly executes thousands of simulations in parallel. PBMO^TM has a capacity of executing 5,000 or more model simulations in parallel if needed. This newly extended method and computational power provide scalable and unsurpassed speed for determining optimal solutions. Hence, the integrated parallel simulation and optimization methods use the breadth and depth search approach; iteratively producing sets candidate investigation points evaluated concurrently instead of the previous and widely used sequential approach (i.e. one model evaluation at a time). Globally optimal solutions are regularly found 100 to 1,000 times or more (wall-clock time) faster than other currently available methods. HGL's optimization approach has also solved highly non-linear problems with over 4,000 decision variables and over 2 million constraints, for which a comparative approach is unknown.

Optimal Design Approach

PBMO^TM project design involves selecting the appropriate combination of its fully comprehensive solution options, including the full capability of mixing and matching various approaches.

• Select proper descriptive model or model set that adequately describes the system to be optimized
• Select optimization objective (e.g. minimize time or cost, maximize mass removal or destruction) and constraints (e.g. maximum flow rates, maximum total and annual funding limits)
• Select computation operating system (MS Windows or Linux) and environment (PC, local area network or Cloud).

The PBMO^TM process formally integrates the stakeholder input into the analysis and provides not only the mathematical "optimal solution" but also "what-if" capability that includes management override control on remedial design analysis. This approach integrates these elements into a single decision framework that is tractable, traceable, and defensible. This solution methodology represents a significant improvement over the non-optimal “trial and error” approach to developing environmental response(s).

METHODS

Understanding optimization rationale and theory, methodological strength, applicability, and limitations, is critical for understanding how to successfully deploy computational optimization analysis on projects. PBMO^TM uses a newly developed breadth and depth search strategy with adaptive partitioning. The breadth search identifies good potential solutions quickly, whereas the depth (local) search drills down to find the optimal solution.

EXAMPLES

Three examples are provided that demonstrate the effectiveness of the optimization method by documenting PBMO^TMs ability to recreate an optimal solution faster than previous methods, ability to solve problems with thousands of decision variables and millions of constraints and to solve highly non-linear flow systems with deep drawdown due to mine dewatering, respectively.

Example 1: Fort Ord, CA

Fort Ord is a former U.S. Army base encompassing more than 28,000 acres in Monterey County, California. Closed by Congress in 1994, the base has been undergoing cleanup with oversight by federal and state regulatory agencies. In December 2003, HGL was contracted by the U.S. Army Corps of Engineers to conduct a Performance-Based Remediation with the objective of Site Closure at Operable Unit (OU)-1. HGL's optimized site exit strategy achieved remedial-action completion in 2014 and obtained regulatory concurrence in 2016, ahead of schedule and below estimated cost. This project won the Grand Prize in the small projects category (AAEES 2017b).

Example 2: Superfund Site, NE

PBMOâ„¢ was used to develop an optimal remedy design for a large chlorinated solvent plume. The problem involved minimizing total life cycle costs, including those for the initial remedy installation, operation and maintenance, and sampling. This problem was enormous and complex, with 4,514 decision variables and over 2 million constraints. A flow and transport model with over one million grid cells was used to describe the contaminant migration through a complex aquifer system. PBMOâ„¢ was used to devise a modified set of constraints to identify a feasible region for optimization. With the final set of mutually agreed upon constraints, the optimal solution was obtained. This effort represents one of the most challenging problems ever solved with an optimization software. There exists no other known problem of this size and complexity from which to supply benchmark comparison.

Example 3: Iron Ore Mine, Western Australia

PBMOâ„¢ was used to design a dewatering strategy for two adjacent iron-ore lenses at an open- pit mine in Australia. The problem was complex, with 113 decision variables and 178 constraints. Additionally, the model used was a non-linear variably saturated flow model with over 200,000 grid cells. PBMOâ„¢ yielded an optimal solution with estimated savings in well installation costs ($4.5M) and water extraction costs ($4.6M).

CONCLUSIONS

The newly massively parallelized PBMO^TM methodology is tested, verified, validated and documented using a broad range of industrial projects. Those results indicate that it meets or exceeds the best available solution either in use, proposed by the subject matter experts or developed by other optimization methods. This new parallelization method now allows for the optimization of remedial designs for the more complicated problems which include thousands of decision variables and millions of constraints. These are precisely the environmental problems that remain in the environmental contamination site portfolio as persistent challenges. The optimal design and implementation of environmental challenge response are now computationally tractable, practical and viable.

REFERENCES

American Academy of Environmental Engineers and Scientists (AAEES, 2017a). 2017 Excellence in Environmental Engineering and Scienceâ„¢ Competition Winner " E3S Grand Prize in Research. Physics-Based Management Optimization Technology for Supporting Environmental and Water Resource Management. Deschaine LM and Guvanasen V. http://www.aaees.org/e3scompetition-winners-2017gp-research.php

(AAEES, 2017b) __ibid. Innovative Approach for Implementing Performance-Based Remediation Projects (http://www.aaees.org/e3scompetition-winners-2017gp-smallprojects.php

Deschaine, L. M. (2016) Numerical Simulation and Optimization of Environmental Treatment Design Using Parallel Computational Methods: Theory and Application. AIChE Annual Meeting, November 13 - 18, San Francisco. https://aiche.confex.com/aiche/2016/webprogram/Paper471994.html

Deschaine, L. M. (2014). Decision support for complex planning challenges. Ph.D. Dissertation, Chalmers University of Technology, Göteborg, Sweden. 233p.

Deschaine, L. M., Lillys, T. P., & Pintér, J. D. (2013). Groundwater remediation design using physics-based flow, transport, and optimization technologies. Environmental Systems Research, 2(1), 1-21. http://www.environmentalsystemsresearch.com/content/pdf/2193-2697-2-6.pdf

Deschaine LM. (2003) Simulation and Optimization of Large Scale Subsurface Environmental Impacts; Investigations, Remedial Design, and Long Term Monitoring. Journal of Mathematical Machines and Systems, Kiev. Number 3,4. Pages 201-218.