Upload the corpora you want to build your engine from. These are typically TMX files, but you can also use TBX, CSV, TSV, or even translated CAT tool files with confirmed segments.
Create a new engine and choose which corpora you want to include.
Hit the Train button. When training finishes, the engine is ready for translation.
A good rule of thumb is to collect at least 100,000 segment pairs for the engine. If the volume is below 100,000 segment pairs, try adding some other resources. Even if they are not 100% relevant to your engine, they can still elevate the overall language quality.
The training process is the step when the machine learns from the corpora. This can happen when you create a new engine, but you can also retrain your existing engines after certain milestones, for example when a project has finished. The process is similar to the continuous improvement of Translation Memories: with every segment added to your engine, it will produce better results.