COMPUtational Framework for Modern Calibration and Validation of Mathematical Models of Subsurface Flows - COMPU-FLOW

预测地下环境的行为（例如，地下水流动和污染物在地下水中的迁移）受到惊人的不确定性的影响。后者主要是因为地下是高度不均匀的，几乎不可能描述其所有细节。非均匀性的分辨率可以通过新类型的实验数据来提高，剩余的不确定性可以通过对状态变量的观测数据进行流/输运模型的校准来降低。不确定性只能通过随机/概率逆建模技术来量化，而不是传统的模型校准方案。大量的各种（随机）逆方法可在文献中。然而，仍然没有对它们的相对优点和缺点作出结论性和令人信服的评估。这一事实为所有当前和未来寻求进一步改进逆建模的研究工作创造了一个具有挑战性的障碍。其中一个关键原因是缺乏定义明确的基准情景，可以在标准化、受控和可重复的条件下对不同的方法进行比较。该提案旨在通过定义一套具有高度准确的参考解决方案的基准情景来克服这一问题。还计划根据这些基准和参考解决方案进行一项全社区范围的比较研究。将长期向研究界提供基准情景、参考解决方案和比较解决方案，供今后继续使用。开发的基准案例将考虑完全饱和的瞬态地下水流，低和高的空间变异性和多高斯以及非多高斯水力传导率场。将特别致力于为基准案例计算高度准确的参考解决方案。参考解决方案将使用本项目期间开发的高度专业化算法生成。该算法将基于马尔可夫链蒙特卡罗的预处理Crank-Nicholson变体，配备自适应建议分布，多回火并行链，随机版本的梯度搜索和非多高斯分布的扩展。参考解决方案将在Jülich的高性能超级计算基础设施上进行计算，然后将开发的算法用于大规模并行计算。地下水逆模型社区将在研讨会上开会，以最终确定比较研究的战略，其中包括物流和性能标准定义等重要主题。共有12个国际知名的团体已经承诺参加他们的逆方法在研讨会和比较研究。总而言之，该提案构成了一项独特的努力，将国际地下水反演建模界聚集在一起，为现有方法提供关键见解并加以改进。

Predicting the behavior of subsurface environments (e.g., groundwater flow and contaminant transport in groundwater) is subject to staggering uncertainties. The latter mainly arise because the subsurface is highly heterogeneous, and it is virtually impossible to characterize all of its details. The resolution of heterogeneity can be improved through new types of experimental data, and the remaining uncertainties may be reduced by calibration of flow/transport models on observed data of state variables. Uncertainty can only be quantified via stochastic/probabilistic inverse modeling techniques instead of conventional model calibration schemes. A large variety of (stochastic) inverse methods is available in the literature. However, a conclusive and convincing assessment of their relative merits and drawbacks is still missing. This fact creates a challenging barrier to all current and future research efforts that seek to further improve inverse modeling. A key reason for this is the lack of well-defined benchmark scenarios against which diverse methods can be compared under standardized, controlled and reproducible conditions. This proposal aims at overcoming this issue by defining a set of benchmark scenarios with highly accurate reference solutions. A community-wide comparison study based on these benchmarks and reference solutions is also planned. Benchmark scenarios, reference solutions and compared solutions will be made available to the research community on a long-term basis for continued future use. The developed benchmark cases will consider fully-saturated transient groundwater flow, low and high spatial variability and multi-Gaussian as well as non-multi-Gaussian hydraulic conductivity fields. Special dedication will be paid to calculating highly accurate reference solutions for the benchmark cases. The reference solutions will be produced with highly specialized algorithms developed during this project. The algorithms will be grounded on the preconditioned Crank-Nicholson variant of Markov Chain Monte Carlo, equipped with adaptive proposal distributions, multi-tempered parallel chains, a randomized version of gradient search and an extension for non-multi-Gaussian distributions. The reference solutions will be calculated on the high performance supercomputing infrastructure in Jülich after adapting the developed algorithms for massive parallel computation. The groundwater inverse modeling community will meet in a workshop to finalize the strategy for the comparison study, which includes important topics like logistics and definition of performance criteria. A total of 12 internationally renowned groups have already committed to participate with their inverse methods in the workshop and comparison study. Altogether, this proposal constitutes a unique effort to bring the international groundwater inverse modeling community together, provide critical insights on existing methods and improve them.