New Methods For Three-dimensional Reconstruction In The Transmission Electron Microscope

Electron tomography is the most popular three-dimensional reconstruction technique in the transmission electron microscopy.It is realized by reconstruction from a series of linear projections of the materials at different tilt angles.However,the massive demand of experimental data complicates the actual problem and forces the technique to face some extra experimental and theoretical problems.Among them,missing wedge,caused by incomplete specimen tilt range,is a vital problem that introduces severe artifacts in the tomogram.On the other hand,as the beam probe size of the scanning transmission electron microscope is reduced,the high-angle annular dark field images will become optical sections,deviating from the linear projections.In this condition,new problems such that whether tomography is theoretically feasible and what the difference is from the reconstruction of bulk materials at low magnification,need to be addressed.In addition,some researchers have proposed several new methods to skip over the negative factors occurring in the redundant images collection process.By combining with some theoretical or simulation results,these methods are capable of analysing the materials' three-dimensional structures from several transmission electron micrographs.Nevertheless,these new techniques still face some theoretical problems and are worth exploring and improving.Focusing on the scientific issues mentioned above,the present dissertation conducts researches through theoretical study,programming simulation,experimental verification and other means,and some innovative results are mainly achieved:(1)A neural network tomography algorithm is developed.It performs the recontruction directly from the experimental data by solving the projection linear equations by a back-propagation neural network model.Different from the regularized or constrained algebraic reconstruction techniques,this algorithm takes advantages of the high complexity of the neural network to minimize missing wedge artifacts and retrieve the missing frequency information without the need for prior knowledge.In addition,the smooth solution resulting from the current algorithm is less susceptible to noise than the results of the general regularization methods as a special average scheme is used in the algorithm.Simulation tests demonstrate that the algorithm can be used to reconstruct various materials with different morphologies.Experimental results show that the algorithm has great advantages over other approaches when the shape of the material is complex and there is experimental noise.(2)Through theoretical simulation,the feasibility of atomic-resolution tomography under nano-scale depth of field is explored.In this study,tilt series of high-angle annular dark field images of atomic models are simulated by multislice method and reconstructed by simultaneous iterative reconstruction technique.Disordered atomic models are used in the study to avoid channelling effect.The influences of the acceleration voltage,semiconvergence angle,defocus,sample's size and element on the reconstruction are explored.The study finds that when the depth of field is less than the sample' thickness,tomography can only correctly reconstruct the local area of the sample.In addition,the actual correct reconstruction area is higher than the nominal focus position of the electron beam,that is,there is a pre-focusing phenomenon.On the other hand,the relative positions between the beam probe and the tilt axis determines the specific positions in the sample,from which the information collected in the images originates,thereby determining the quality of the reconstruction.When the electron beam probe is focused on the tilt axis,the best reconstruction quality can be obtained.Otherwise,it is equivalent to including missing wedge in the collected information.(3)A feasible three-dimensional reconstruction technique for reconstructing the surface of general crystals,through quantitative analysis of a single atomic transmission electron micrograph,is proposed.Firstly,a global matching algorithm based on quantitative simulation analysis is developed.By detailed simulation tests,the necessity of the global matching algorithm is proved,for accurate reconstrction from a general atomic transmission electron micrograph.In addition,the technique adopts a self-validation scheme,taking an average result of multiple independent reconstructions as the final result.The resolution of the three-dimensional reconstruction can be estimated from the multiple results.Acorrding to the result and simulation analysis,a confidence factor is introduced to quantitatively explore the influence of amorphous on the reconstruction result.Applying the proposed algorithm to a two-dimensional experimental image from a Si[110] crystal sample,it is shown that an atomic-resolution of one interatomic distance(= 0.384 nm)in three-dimension for both the height(defocus)and the thickness(atom numbers)of Si atomic columns can be achieved,provided that the covering amorphous layers were less than 1.0 nm in thickness.This dissertation proposes two new techniques to realise high precision three-dimensional reconstruction,and reveals some new theorectical problems that will be faced in atomicresolution electron tomography.These new theories and techniques have a lot of potential for better supporting material research.