Objective: To significantly improve the positive predictive value (PPV) and sensitivity of Applied Biosystems™ Axiom™ array variant calling, by means of novel improvement to genotyping algorithms and careful quality control of array probesets. The improvement makes array genotyping more suitable for very rare variants. Design: Retrospective evaluation of UK Biobank array data re-genotyped with improved algorithms for rare variants. Participant: 488,359 people recruited to the UK Biobank with Axiom array genotyping data including 200,630 with exome sequencing data. Main Outcome Measures: A comparison of genotyping calls from array data to genotyping calls on a subset of variants with exome sequencing data. Results: Axiom genotyping [18] performed well, based on comparison to sequencing data, for over 100,000 common variants directly genotyped on the Axiom UK Biobank array and also exome sequenced by the UK Biobank Exome Sequencing Consortium. However, in a comparison to the initial exome sequencing results of the first 50K individuals, Weedon et al. [1] observed that when grouping these variants by the minor allele frequency (MAF) observed in UK Biobank, the concordance with sequencing and resulting positive predictive value (PPV) decreased with the number of heterozygous (Het) array calls per variant. An improved genotyping algorithm, Rare Heterozygous Adjustment (RHA) [16], released mid-2020 for genotyping on Axiom arrays, significantly improves PPV in all MAF ranges for the 50K data as well as when compared to the exome sequencing of 200K individuals, released after Weedon et al. [1] performed their comparison. The RHA algorithm improved PPVs in the 200K data in the lowest three frequency groups [0, 0.001%), [0.001%, 0.005%) and [0.005%, 0.01%) to 83%, 82% and 88%; respectively. PPV was above 95% for higher MAF ranges without algorithm improvement. PPVs are somewhat higher in the 200K dataset, due to a different "truth set" from exome sequencing and because monomorphic exome loci are not included in the joint genotyping calls for the 200K data set, as explained in the methods section. Sensitivity was higher in the 200K data set than in the original 50K data as well, especially for low MAF ranges. This increase is in part due to the larger data set over which sensitivity could be computed and in part due to the different WES algorithms used for the 200K data [7]. Filtering of a relatively small number of non-performing probesets (determined without reference to the exome sequencing data) significantly improved sensitivities for all MAF ranges, resulting in 70%, 88% and 94% respectively in the three lowest MAF ranges and greater than 98% and 99.9% for the two higher MAF ranges ([0.01%, 1%), [1%, 50%]). Conclusions: Improved algorithms for genotyping along with enhanced quality control of array probesets, significantly improve the positive predictive value and the sensitivity of array data, making it suitable for the detection of very rare variants. The probeset filtering methods developed have resulted in better probe designs for arrays and the new genotyping algorithm is part of the standard algorithm for all Axiom arrays since early 2020.