Assessment of Pile Design Methods Using Advanced Data AnalyticsΒΆ

PhD Dissertation by Nikolaos Machairas

Important

No license, yet. You may not use or reference any of the material presented herein. For all inquiries, please contact me.

For the past 30+ years, engineers and researchers have been independently collecting pile load tests and relevant subsurface data while organizing this information into structurally dissimilar repositories. Despite their competent efforts, the overall result was highly fragmented with very little benefit to the greater geotechnical community. Meanwhile, scientists aided by state-of-the-art data analytics have been transforming their respective industries, producing remarkable predictions and insights. The current unstructured and decentralized scheme of valuable pile load test data has provided few benefits. Instead it has been a hindrance to the geotechnical community at large.

Use of load test databases for comparison between calculated and interpreted capacities has provided insights on the suitability of use of design methods under varying pile and soil conditions. Past studies have generally demonstrated that all methods in use for calculating the ultimate capacity of single piles have large margins of error. This dissertation verified past findings and expanded the evaluation for Large Diameter Open Ended Piles (LDOEP) discovering similarly poor performance.

As part of this doctoral dissertation, a multi-tiered system, called NYU Pile Capacity was developed that allowed for the collaborative data storage, cleaning and analysis of data for deep foundations. NYU Pile Capacity has a relational database backend and a friendly HTML interface for user interactions. Existing load test databases have been delivered as locally installed software applications. NYU Pile Capacity, however, required no software installations and was served over the Internet as a web application. Users can log in and instantly start running custom aggregate analyses on the 5,000+ records that were imported from existing datasets or add new records.

NYU Pile Capacity was built using Python Flask and can batch-process multiple load test records for practically countless combinations of soil conditions and pile types. Furthermore, NYU Pile Capacity was designed to be extended in order to run additional analyses with minimal updates to its core codebase.

Most of the methods in current use for pile design are based on empirical formulas that required gross overgeneralization to develop. The empirical/semi-empirical design guidelines were derived from as few as 41 load test records. This doctoral dissertation compiled a dataset of more than 5,000 load test records and evaluated popular methods for capacity calculation and capacity interpretation against this massive dataset. The results of analyses revealed that contrary to common practice and against federal and state guidelines, the recommended Nordlund and Tomlinson methods were not producing optimal designs. Instead, the less popular API and Lambda methods proved far superior. Also, a comprehensive evaluation of interpreted capacity methods validated the dominance of the original Davisson method while not finding any significant benefits to the subsequent federally proposed modifications to this method, once again proving that going against Federal guidelines could produce more efficient designs.

Another major finding of this dissertation is that given a large enough training sample, pile capacity can be reliably estimated by employing Machine Learning techniques. In projects that involve a large number of pile foundations, not all piles are individually designed and checked. Existing design software do not run aggregate analyses, and manually repeating the process hundreds of times would be extremely time consuming. However, working off of a reliable approximation of subsurface conditions for the entire site based on the results of site investigation, every pile on site can be designed or checked via batch processing and an iterative optimization process. A proof-of-concept of this alternative design process was presented where a Support Vector Machine algorithm outperformed the Federal design method for driven piles. Perhaps more remarkably, the predictive model outperformed the FHWA pile design method by relying only on seven readily available features as compared to a laborious and error-prone design methodology.

Finally, this dissertation presented the argument for the case-based design of driven pile foundations. Most of the existing design methods attempted to generalize and provide recommendations for all soil conditions and all pile types. There was, however, little focus on the performance of these methods for specific soil conditions and pile types. The industry implemented the design methods as blanket solutions expecting that they would perform well for all cases. The custom tools developed in this study provided the flexibility to run aggregate analyses on groups of load test records with similar characteristics. The results of these analyses revealed that design methods do not perform equally well for all cases. Gaining insights into the cases for which design methods perform best can enable enhanced pile design workflows where instead of using a single method to calculate capacities, a combination of methods can be employed depending on the soil conditions and pile type. Hence, case-based design can lead to safer and cost-effective designs for driven pile foundations.


Please use the table of contents to navigate.



Table of contents:


Appendices: