This work presents techniques for addressing the black box problem for deep learning in high-energy physics applications at the LHC. In an initial group of studies, a method is presented for translating a black box classifier using high-dimensional detector data into a minimal set of simple physics motivated features with equivalent classification performance. The strategy is first applied to a benchmark discrimination task for jets from a boosted W boson decay. The algorithm is then used on two active areas of standard model research at the LHC: electron identification and prompt muon isolation. Finally, the technique is applied to a beyond the standard model study of semi-visible jets produced via a theoretical dark quark hadronization process.

A second technique is presented, providing a method for the explicit embedding of physics parameters alongside measured features in a deep learning model. This architecture yields a parameterized classifier that can smoothly interpolate between physical features. The result is a simpler and more powerful machine learning classifier that can seamlessly incorporate expert physics knowledge into its learned solution. The parameterized network is applied to a benchmark classification task for tt ̄ decays.

Several theoretical parameter spaces are analysed using techniques from machine learning. First, machine learning is used to leverage high dimensional detector information to constrain the mass and coupling of a simplified model which extends the standard model with a heavy Z' boson. The high dimensional information is seen to improve the exclusion contours of the analysis relative to a low dimensional analysis that instead uses summary statistics. Next, generative models are used to sample a high dimensional parameter space of a supersymmetric model subject to a constraint. Generative models are seen to provide an increase in efficiency of over an order of magnitude in a search for models that satisfy the constraint when compared to random sampling. Finally, we analyse a dataset of supersymmetric theories to understand how they may best discovered in experiment. We propose a set of experiments to be performed at the Large Hadron Collider that are most sensitive to these models.

This thesis covers a new search for dark matter at the LHC in addition to improved particle identification techniques using neural networks. Models are introduced for a new search at the LHC, mono-Z', where dark matter is produced in association with a new hypothetical boson. Searches are performed using data collected from p p collisions at 13 TeV using the ATLAS detector. Measurements from this search are consistent with the Standard Model, and upper limits are estimated on the production cross sections and model parameters. Next, we explore techniques for identification of for jets and muons using deep convolutional neural networks trained on calorimeter energy deposits. In both studies, improvements on the baseline techniques using high-level engineered features is observed.

We demonstrate techniques to improve the performance of data driven methods used in collider experiments, through the use of neural networks. First, using a simulated muon dataset, we probe the discriminating power of the typically used isolation observable by comparing it to neural networks trained on full event details. By performing a search over the space of Energy Flow Polynomials (EFPs), a set of scalar observables which performs similarly to the full information is identified. This methodology is then applied to real collider data obtained from CMS Open Data. The CMS data lacks event level class labels, necessitating the use of CWoLa, a weakly supervised training method, along with an sPlotsbased performance evaluation method. Once again, we successfully identify a minimal set of scalar observables capable of outperforming isolation. Finally, a novel data-augmentation scheme is introduced. Symmetries present in an ideal dataset may be broken by detector effects, leading to lower quality augmented copies. We perform augmentation pre-detection in simulation, and further encourage invariance across augmented copies during training. We find that synthesizing examples this way leads to faster convergence, and that encouraging invariance yields further performance gains.

This work explores the application of machine learning for multi-prong jet classification at the LHC. We compare the performance of high-level networks trained on jet observables to low-level networks trained on the full event information. Our focus is on an extreme scenario, evaluating the performance of the classifiers on jets with a large number of sub-jets, larger than previously tested in the literature. The results indicate that traditional jet observables like $N$-subjettiness and Energy-Flow Polynomials are good discriminants when tested on jets with up to eight hard sub-jets, but that there is information in the events that is untapped by these observables. We introduce Jet Rotational Metrics, a new family of observables designed to capture features of the degree of discrete rotational symmetry of jets. These observables prove highly valuable for the classification task, bridging the gap between the low- and high-level networks.

This dissertation also introduces a technique for the accurate estimation of systematic uncertainties in physics studies using Gaussian process regression. A typical approach is to assume the factorization of the various sources of systematic uncertainties. This approach is often extended to assume that the impact of these individual sources of uncertainty on observables of the detector response also factories. Our technique uses Gaussian processes to model observables as functions of the nuisance parameters. We show that this technique is more accurate than the factorized approach and that it can learn from limited samples by including gradient information. Additionally, we present a Bayesian-based sampling strategy that efficiently explores the space of experimental response while reducing the predictive uncertainty of the Gaussian process.

A search for new physics resonances in the dijet invariant mass spectrum is presented here. Dijet events are collected at center of mass energy of 13 TeV with the ATLAS detector at the Large Hadron Collider in 2015 and 2016, equating to a total integrated luminosity of 37 $fb^{-1}$. This data is compared to background predictions, and no significant deviations from the expected is seen. Therefore, the dataset is used to set improved upper limits on the mass of four benchmark signal models and one generic model at 95\% CL. These limits exclude excited quarks with masses below 6.0 TeV, quantum black holes below 8.9 TeV, heavy W' boson masses below 3.6 TeV, and $W^{*}$ bosons masses below 3.4 TeV and between 3.77-3.85 TeV; as well as limits on a range of masses and couplings in a $Z'$ dark matter mediator model. Model-independent limits are also set on signals with a Gaussian shape at various mass resolutions. Finally, a proof of concept study is done on a new method to predict dijet backgrounds, which may be implemented in future analyses.

In this work, several phenomenological models for extending the Standard Model are presented, with an emphasis on providing a candidate dark matter particle.

These models are specifically designed as benchmark signals in two novel search channels proposed for LHC experiments: mono-$Z$ and mono-Higgs.

Searches for dark matter production in these newly-proposed channels are then conducted using data from $pp$ collisions at $\sqrt{s} = 8 \TeV$ recorded by the ATLAS detector.

In both cases, the observed measurements are consistent with the predicted Standard Model rates.

Therefore, upper limits are derived on the production cross sections of these dark matter models, which are also translated to limits on theoretical parameters specific to each model.

Lastly, a method of searching for hadronic resonances with masses in the 20--500 \GeV\ range at the LHC is described in the context of a minimal leptophobic $Z'$ extension to the Standard Model.

The feasibility of this search method is investigated assuming a dataset of 4 \ifb\ of $pp$ collisions in terms of the sensitivity to the $Z'$ coupling parameter.

These searches are expected to provide nontrivial limits on the $Z'$ model in regions that have been either only weakly constrained or never probed by direct measurements.

This dissertation examines the feasibility of appropriating the global network of smartphones as an observatory for ultra-high-energy cosmic rays. An application and cloud-based data-acquisition system are first proposed for such an observatory, in which heterogenous devices self-calibrate and server feedback optimizes individual device triggers. Methods of monitoring and manually controlling this network are also examined. Detection efficiencies for cosmic muons, MeV-scale gamma rays, and 120 GeV protons are then measured for this trigger, with which a Monte Carlo pixel model is fine-tuned. Extending this model to cosmic ray showers, the effective area of the observatory for super-GZK primaries is shown to equal that of the Pierre Auger Observatory with a global participation rate under 0.1%, far below initial estimates presented in Ref. [83]. A vastly improved sensitivity to photons and more precise modeling of electrons are primarily responsible for this discrepancy, though more refined models of the combinatorial background may yet lead to more stringent requirements.

In this thesis a few approaches are explored to advance resonance finding searches in collider physics. Two analyses in the dimuon and dijet initial-state radiation(ISR)final state resonance are searched using proton-proton collisions at √s=13 TeV with an integrated luminosity of 139/fb, collected by the ATLAS experiment. No statistically significant excess is seen in the dijetISR search. In the dimuon search, a novel background estimation method is developed, in which the smooth but unknown background is estimated using a Gaussian process fit.

Until the construction of the aptly-named cosmotron in the early 1950s, particle physicists relied on cosmic ray tracks in photographic emulsions and cloud chambers to discover antimatter and subatomic particles. Nearly 110 years since their discovery, the origin and composition of the highest energy cosmic rays remains largely a complete mystery.

In that time, solid-state pixel technology has become a mainstay in both particle detectors and consumer smartphone cameras, but for largely economic reasons, modern cosmic ray surface detectors are primarily water-Cherenkov or plastic-scintillator type. However, with both the worldwide number of smartphone users exceeding 3 billion and at least as many laptop computers in use, consumer solid-state pixel sensors (cameras) have a combined surface area over 5 times the cross-sectional area of the Pierre Auger Observatory's 1,660 water-Cherenkov detectors.

In this dissertation, I discuss the potential, the process and the problems faced in turning the populated planet into a cosmic ray telescope using smartphone cameras. In Chapter 2, I develop novel extensive air shower longitudinal muon and photon density models that clearly exhibit better agreement with CORSIKA simulations than popular alternatives. I also provide a parameterization scheme that spans variations in primary energy, inclination angle, and observation height. In Chapter 4, I identify muon and photon signatures present in real CRAYFIS user data, propose a novel test array of CRAYFIS-enabled smartphones, and present a high-performance data acquisition application. At last, in Chapter 5, I calculate the sensitivity of a global CRAYFIS network to simultaneous extensive air showers as a function of observation time and incident flux, and find that at least 1 million CRAYFIS users worldwide are needed to identify novel phenomena signal over background at 3$\sigma$ statistical significance over a reasonable time-span.