The project aims to accel­er­ate and improve the fabri­ca­tion of micro­ma­te­ri­als by 3D laser print­ing through the use of deep neural networks (DNNs). Physi­cal simula­tions of the print­ing process are devel­oped and used to train the DNNs. They can then, for example, charac­ter­ize the printed struc­tures already in the printer or pre-compen­sate objects in such a way that itera­tive charac­ter­i­za­tion and optimiza­tion outside the printer can be minimized.

3D laser print­ing is a versa­tile manufac­tur­ing approach for various optical elements or materi­als with special and extreme proper­ties. A struc­ture is printed by focus­ing a laser beam into a photore­sist, which then polymer­izes. Once print­ing is complete, the photore­sist is removed with a solvent, leaving the printed struc­ture behind. Subse­quent charac­ter­i­za­tion usually reveals that the printed object does not look exactly as it was designed. This is the start­ing point for a sometimes weeks-long optimiza­tion of the materi­als to obtain the desired struc­ture after the process.

The topic of the doctor­ate is to simplify and accel­er­ate this optimiza­tion process using various algorithms in the field of deep learn­ing. The deep neural networks (DNNs) are trained to invert simula­tions based on physi­cal equations, i.e. to calcu­late the initial state that led to a certain obser­va­tion. The differ­ent DNNs can then be used to recon­struct objects during and/or after print­ing to reduce tedious charac­ter­i­za­tion outside the printer. Further­more, it will be inves­ti­gated to what extent DNNs can be used to pre-compen­sate objects so that the desired object appears after print­ing, and optimiza­tion by the users is no longer neces­sary or at least minimized.