Accurate forecasts of weather and climate will become increasingly important as the world adapts to anthropogenic climatic change. Forecasts' accuracy is limited by the computer power available to forecast centres, which determines the maximum resolution, ensemble size and complexity of atmospheric models. Furthermore, faster supercomputers are increasingly energy-hungry and unaffordable to run. In this thesis, a new means of making computer simulations more efficient is presented that could lead to more accurate forecasts without increasing computational costs. This 'scale-selective reduced precision' technique builds on previous work that shows that weather models can be run with almost all real numbers represented in 32 bit precision or lower without any impact on forecast accuracy, challenging the paradigm that 64 bits of numerical precision are necessary for sufficiently accurate computations. The observational and model errors inherent in weather and climate simulations, combined with the sensitive dependence on initial conditions of the atmosphere and atmospheric models, renders such high precision unnecessary, especially at small scales. The 'scale-selective' technique introduced here therefore represents smaller, less influential scales of motion with less precision. Experiments are described in which reduced precision is emulated on conventional hardware and applied to three models of increasing complexity. In a three-scale extension of the Lorenz '96 toy model, it is demonstrated that high resolution scale-dependent precision forecasts are more accurate than low resolution high-precision forecasts of a similar computational cost. A spectral model based on the Surface Quasi-Geostrophic Equations is used to determine a power law describing how low precision can be safely reduced as a function of spatial scale; and experiments using four historical test-cases in an open-source version of the real-world Integrated Forecasting System demonstrate that a similar power law holds for the spectral part of this model. It is concluded that the scale-selective approach could be beneficially employed to optimally balance forecast cost and accuracy if utilised on real reduced precision hardware.

Large-scale bed features are often encountered in coastal waters, and include sandbanks and spoil heaps. The morphodynamic development of such features involves complicated nonlinear interactions between the flow hydrodynamics, sediment transport, and bed profile. Numerical modelling of the morphodynamic evolution and migration of large-scale bed features is necessary in order to understand their long-term behaviour in response to changing environmental conditions. This thesis describes detailed measurements of the morphodynamics of sand mounds in unidirectional and oscillatory (tidal) flows, undertaken at the U.K. Coastal Research Facility (UKCRF). High quality data were collected, including water velocities, water levels and overhead images. The parameters tested are: three types of mound shape (circular and elliptical in plan shape, and Gaussian, cosine and triangular in cross-section); underlying fixed or mobile bed conditions; and initial crest height (submerged, surface-touching and surface-piercing). Peak flow velocities are about 0.5 m/s, the sand median grain size is 0.454 mm, and transport occurring mostly as bedload. When analysing the data, the bed contours are determined by digitising the shoreline at different water levels. From these plots, the volume, height, and centroid position of the mound are calculated. A large-scale fit method, based on a Gaussian function has been used to separate small-scale ripples from the large-scale bed structure during the evolution of an isolated sand mound or spoil heap. The bed profile after the ripples are removed is comparable to typical predictions by shallow-flow numerical solvers. The UKCRF experiments investigated the morphodynamic response of a bed mound to hydrodynamic forcing: shape changes, migration rates, volume decay and sediment transport rates. The measured migration rate and decay of a submerged sand mound in the UKCRF are found to be in satisfactory agreement with results from various theoretical models, such as the analytical solution derived by De Vriend. Numerical predictions of mound evolution by a commercial code, PISCES, are also presented for a fully submerged sand mound; the bed evolution is reasonably similar to that observed in the UKCRF. The data provided as a result of the research reported in this thesis provide insight into the behaviour of sand mounds in steady and unsteady flows at laboratory scale, and should also be useful for benchmark (validation) purposes to numerical modellers of large-scale morphodynamics.

In this thesis, the problem of predicting the full dose distribution from a partially modeled dose calculation is addressed. Two solutions were studied: a vanilla Hierarchically Densely Connected U-net (HDUnet) and a Conditional Generative Adversarial Network (CGAN) with HDUnet as a generator. The CGAN approach is a 3D version of Pix2Pix [1] for Image to Image translation which we name Dose2Dose. The research question that this project tackled is whether the Dose2Dose can learn more effective dose transformations than the vanilla HDUnet. To answer this, the models were trained using dose calculations of phantom slabs generated for the problem in pairs of inputs (doses without magnetic field) and targets (doses with magnetic field). Once trained, the models were evaluated and compared in various aspects. The evidence gathered suggests that the vanilla HDUnet model can learn to generate better dose predictions than the generative model. However, in terms of the resulting dose distributions, the samples generated from the Dose2Dose are as likely to belong to the target dose calculation distribution as those of the vanilla HDUnet. The results contain errors of considerable magnitude, and do not accomplish clinical suitability tests. / I denna avhandling har problemet med att förutsäga full dosfördelning från en delvis modellerad dosberäkning tagits upp. Två lösningar studerades: ett vanilla HDUnet och ett betingat generativt nätverk (CGAN) med HDUnet som generator. CGAN -metoden var en 3D-version av Pix2Pix [1] för översättning av bild till bild med namnet Dose2Dose. Forskningsfrågan som detta projekt tog upp var om Dose2Dose kan lära sig mer effektiva dostransformationer än vanilla HDUnet. För att svara på detta tränades modellerna med hjälp av parvisa dosberäkningar, i indata (doser utan magnetfält) och mål (doser med magnetfält).. När de var tränade utvärderades modellerna och jämfördes i olika aspekter. De samlade bevisen tyder på att Vanilla HDUnet -modellen kan lära sig att generera bättre dosförutsägelser än den generativa modellen. När det gäller de resulterande dosfördelningarna är emellertid de prover som genererats från Dose2Dose lika sannolikt att tillhöra måldosberäkningsfördelningen som de för vanilla HDUnet. Resultaten innehåller stora storleksfel och uppfyller inte kraven för klinisk tillämpbarhet.