[
    {
        "id": "thesis:18919",
        "collection": "thesis",
        "collection_id": "18919",
        "cite_using_url": "https://resolver.caltech.edu/CaltechTHESIS:09102026-220059166",
        "primary_object_url": {
            "basename": "bourdais_theo_2027.pdf",
            "content": "final",
            "filesize": 25448608,
            "license": "other",
            "mime_type": "application/pdf",
            "url": "/18919/2/bourdais_theo_2027.pdf",
            "version": "v5.0.0"
        },
        "type": "thesis",
        "title": "Advances in Scientific Discovery Automation",
        "author": [
            {
                "family_name": "Bourdais",
                "given_name": "Th\u00e9o J.",
                "orcid": "0009-0007-9439-2349",
                "clpid": "Bourdais-Theo-J"
            }
        ],
        "thesis_advisor": [
            {
                "family_name": "Owhadi",
                "given_name": "Houman",
                "orcid": "0000-0002-5677-1600",
                "clpid": "Owhadi-H"
            }
        ],
        "thesis_committee": [
            {
                "family_name": "Mazumdar",
                "given_name": "Eric V.",
                "orcid": "0000-0002-1815-269X",
                "clpid": "Mazumdar-Eric"
            },
            {
                "family_name": "Chandrasekaran",
                "given_name": "Venkat",
                "clpid": "Chandrasekaran-V"
            },
            {
                "family_name": "Owhadi",
                "given_name": "Houman",
                "orcid": "0000-0002-5677-1600",
                "clpid": "Owhadi-H"
            },
            {
                "family_name": "Sahai",
                "given_name": "Tuhin",
                "orcid": "0000-0003-1896-8768",
                "clpid": "Sahai-Tuhin"
            }
        ],
        "local_group": [
            {
                "literal": "div_eng"
            }
        ],
        "abstract": "<p>Much of machine learning for scientific computing approximates a function specified in advance. This thesis investigates how dedicated learning systems can address tasks of increasing complexity and thereby automate larger parts of scientific discovery. Each method exploits structure in the learning method or in the scientific problem to constrain the search and incorporate mathematical knowledge.</p>\r\n\r\n<p>In Function Learning, the target task is prescribed; we develop minimal-variance aggregation of black-box predictions, operator learning for partial differential equations at machine precision, and pruning of deep networks guided by random matrix theory.</p>\r\n\r\n<p>In Structure Discovery, the relationships among variables must be recovered from data; we identify functional and stochastic dependency graphs for digital twins with up to one thousand variables while treating randomness, time, and control explicitly.</p>\r\n\r\n<p>In Algorithm Discovery, we design algorithms by combining computational primitives through a variant of AlphaZero and expand the primitive set through Algorithmic Byte Pair Encoding. The resulting algorithms adapt during execution and outperform state-of-the-art methods on the strongly NP-hard Quadratic Assignment Problem; the framework also produces a new formulation of Grover's algorithm using half as many gates as the standard construction.</p>",
        "doi": "10.7907/m43v-6e69",
        "publication_date": "2027",
        "thesis_type": "phd",
        "thesis_year": "2027"
    }
]