Finding new peptides that can bind to some inorganic materials with high affinity and specificity is an active research problem with many application areas ranging from medicine to electronics. While it is easier to design peptides in-silico, in-vitro testing of binding is a slow and expensive process. Peptides with similar structure usually have similar functions. Score matrices, such as Blosum, PAM and Gonnet are used to score amino acid replacements in alignment algorithms such as Needleman-Wunsch or Smith-Waterman which measure similarity between two sequences. These score matrices may not represent the frequency of amino-acids replacement for experimental data. Score matrices which are optimized for specific peptide class or function have an important effect on accurate classification of artificial or newly discovered peptides. In this paper, score matrices that are specific to quartz binding peptides are derived using an evolutionary strategy (ES) and performance of these matrices are compared with the Quartz I matrix of Oren et.al. (2007). Results show that the matrices which obtained from the ES have better performance than Quartz I matrix when TSS (total similarity score) is used as a performance measurement criterion. Using ES scoring matrices and TSS for classifying peptides as strong, moderate or weak has performed 30% better in terms of accuracy and 40% better in terms of the number of inversions of affinity ordering of peptides than Quartz I matrix.