In-line　terahertz　pulsed　imaging　has　been　utilized　to　measure　the　film　coating　thickness　of　individual　tablets　during　the　coating　process　in　a　production-scale　pan　coater.　A　criteria-based　waveform　selection　algorithm　(WSA)　was　developed　to　select　terahertz　signals　reflected　from　the　surface　of　coating　tablets　and　determine　the　coating　thickness.　Since　the　WSA　uses　many　criteria　thresholds　to　select　terahertz　waveforms　of　sufficiently　high　quality,　it　could　reject　some　potential　candidate　tablet　waveforms　that　are　close　but　do　not　reach　the　threshold　boundary.　On　the　premise　of　the　availability　of　large　datasets,　we　aim　to　improve　the　efficiency　of　WSA　with　machine　learning.　This　article　presents　a　recurrent　neural　network　approach　to　optimize　waveform　selection.　In　comparison　with　the　conventional　method　of　WSA,　our　approach　allows　more　than　double　the　number　of　waveforms　to　be　selected　while　maintaining　excellent　agreement　with　offline　thickness　measurements.　Moreover,　the　processing　time　of　waveform　selection　decreases　so　that　it　can　be　applied　for　real-time　coating　monitoring　in　the　pharmaceutical　industry,　which　leads　to　more　advancements　in　quality　control　for　pharmaceutical　film　coating.